Merge pull request 'Feature/tts voice management' (#105) from feature/tts-voice-management into master

Reviewed-on: #105

.gitignore

@@ -5,6 +5,8 @@ database/target
 *.db-shm
 *.db-wal
 .env
+# Server-local TTS pronunciation overrides (tts_pronunciations.example.json is the template)
+/tts_pronunciations.json
 /tmp
 /docs
 /specs

CLAUDE.md

@@ -645,6 +645,14 @@ OPENROUTER_APP_TITLE=ImageApi                  # Optional attribution header
 # re-embedding — mixed vector spaces break similarity search.
 LLM_BACKEND=ollama
 
+# Embedding model contract. Corpus and queries must be embedded by the same
+# model with matching prefixes — after changing the embed model or any of
+# these, run `cargo run --bin reembed_embeddings` (all tables) or search is
+# garbage. Prefix values may contain a literal \n (expanded to a newline).
+EMBEDDING_DIM=768           # 768 = nomic-embed-text v1.5; 1024 = Qwen3-Embedding-0.6B
+EMBED_QUERY_PREFIX=         # nomic: "search_query: " | Qwen3: "Instruct: <task>\nQuery: "
+EMBED_DOCUMENT_PREFIX=      # nomic: "search_document: " | Qwen3: leave empty
+
 # llama.cpp / llama-swap (used when LLM_BACKEND=llamacpp). OpenAI-compatible
 # proxy hosting one or more llama-server processes. Chat models receive
 # images directly via content-parts (all models assumed vision-capable).
@@ -668,6 +676,8 @@ LLAMA_SWAP_TTS_REF_SECONDS=30                  # Max voice-clone reference clip
                                                # (Chatterbox is zero-shot; ~10-20s clean ref is ideal)
 LLAMA_SWAP_TTS_REQUEST_TIMEOUT_SECONDS=600     # Per-request synth timeout (long chunked insights take
                                                # minutes); overrides the shared client timeout for /tts/speech
+TTS_PRONUNCIATIONS_PATH=tts_pronunciations.json # JSON map of pronunciation overrides applied before synth
+                                               # (see tts_pronunciations.example.json); hot-reloaded on change
 
 # Insight Chat Continuation
 AGENTIC_CHAT_MAX_ITERATIONS=6                  # Cap on tool-calling iterations per chat turn (default 6)

README.md

@@ -153,17 +153,39 @@ behind the same llama-swap proxy. Only requires `LLAMA_SWAP_URL` (the TTS client
 is built whenever that's set — independent of `LLM_BACKEND`). Endpoints:
 - `POST /tts/speech` — body `{ text, voice?, format?, exaggeration?, cfg_weight?,
   temperature? }`; returns `{ audio_base64, format }`. Input is cleaned
-  server-side (markdown + emoji stripped) and the generation knobs are clamped
+  server-side (markdown + emoji stripped, then pronunciation overrides applied —
+  see below) and the generation knobs are clamped
   to Chatterbox's ranges. Synthesis is serialized (one at a time — the upstream
   has no GPU lock of its own); a concurrent request gets a fast `429`.
-- `GET /tts/voices` — list the voice library.
+- `POST /tts/speech/jobs` — durable variant for long syntheses: same body as
+  `/tts/speech`, returns `202 { job_id, status }` immediately. Jobs queue on the
+  GPU permit instead of fast-failing `429`.
+- `GET /tts/speech/jobs/{id}` — poll a job: `{ job_id, status, format,
+  audio_base64?, error? }` with status `queued|running|done|error|cancelled`.
+  Results are kept in memory ~10 min after completion, then the job 404s.
+- `DELETE /tts/speech/jobs/{id}` — cancel a queued/running job.
+- `GET /tts/voices` — list the voice library. Served from an in-memory cache
+  (so the listing doesn't make llama-swap spin up the TTS model and evict the
+  resident LLM); pass `?refresh=1` to force an upstream re-query. The cache is
+  invalidated by voice create/delete.
 - `POST /tts/voices/upload` — multipart `voice_name` + `voice_file`; clone a
   voice from an uploaded clip (≤25 MB).
 - `POST /tts/voices/from-library` — body `{ voice_name, path, library? }`; clone
   from a library file (audio forwarded as-is; video has its audio extracted via
   ffmpeg).
+- `DELETE /tts/voices/{name}` — remove a cloned voice from the library.
+
+Created voice names are tagged with the ref-clip cap in effect (e.g.
+`grandma-30s`) so the library shows which reference length produced each clone.
+
+Words the model mispronounces (place names, initialisms) can be rewritten
+before synthesis via a JSON map — copy `tts_pronunciations.example.json` to
+`tts_pronunciations.json` and edit; changes apply without a restart. Full
+matching rules are documented in `src/ai/pronunciation.rs`.
 
 Env:
+- `TTS_PRONUNCIATIONS_PATH` - pronunciation-override JSON file
+  [default: `tts_pronunciations.json` in the working directory]
 - `LLAMA_SWAP_TTS_MODEL` - TTS model id in llama-swap's `config.yaml` [default: `chatterbox`]
 - `LLAMA_SWAP_TTS_VOICE` - default voice used when a `/tts/speech` request omits `voice` (optional)
 - `LLAMA_SWAP_TTS_REF_SECONDS` - max voice-clone reference clip length in seconds

src/ai/gpu.rs

@@ -0,0 +1,88 @@
+// GPU lease — in-process coordination for llama-swap model contention.
+//
+// llama-swap runs the heavyweight models (chat / vision / Chatterbox TTS) as
+// a mutually-exclusive set on one GPU (matrix DSL `(q27 | … | tts) & e`): a
+// request for a non-resident model is HELD by llama-swap until the resident
+// model's in-flight requests drain, then the models swap. That hold counts
+// against the *holder's* reqwest timeout — measured live: a queued TTS burned
+// 77s of its budget behind a single LLM turn, and an LLM request behind a
+// running synthesis waited the entire remaining synth. Uncoordinated
+// cross-model traffic therefore times out instead of queueing.
+//
+// The lease moves that wait into this process, BEFORE the HTTP request is
+// sent and before its timeout starts:
+// - chat/vision requests (the LLM-side slots) share the READ lease;
+// - TTS synthesis and voice-library ops (anything that spins Chatterbox up
+//   and evicts the LLM) take the WRITE lease;
+// - embeddings take NO lease: the `embed` slot is in llama-swap's
+//   always-resident group (the `& e` term) and never participates in a swap,
+//   so leasing it would only stall searches behind a queued synthesis.
+//
+// tokio's RwLock is fair (FIFO, write-preferring): a queued TTS gets the GPU
+// right after the current LLM request drains, and later LLM requests queue
+// behind it — bounded waits in both directions, no starvation, no timeout
+// budget burned while waiting.
+//
+// RULES: hold a lease for exactly one HTTP request (for streaming, the
+// stream's lifetime) and NEVER acquire one while already holding one — once a
+// writer is queued, new read acquisitions block, so nested acquisition can
+// deadlock.
+
+use std::sync::LazyLock;
+use std::time::Instant;
+use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
+
+static GPU_LEASE: LazyLock<RwLock<()>> = LazyLock::new(|| RwLock::new(()));
+
+/// Waits longer than this are logged — they mean a cross-model swap was
+/// avoided and quantify what the request *would* have burned of its timeout.
+const SLOW_WAIT_LOG_SECS: f64 = 2.0;
+
+/// Shared lease for LLM-side requests (chat / vision slots).
+pub async fn llm_lease() -> RwLockReadGuard<'static, ()> {
+    let started = Instant::now();
+    let guard = GPU_LEASE.read().await;
+    log_slow_wait("llm", started);
+    guard
+}
+
+/// Exclusive lease for TTS-side requests (speech synthesis + voice-library
+/// ops that spin up Chatterbox).
+pub async fn tts_lease() -> RwLockWriteGuard<'static, ()> {
+    let started = Instant::now();
+    let guard = GPU_LEASE.write().await;
+    log_slow_wait("tts", started);
+    guard
+}
+
+fn log_slow_wait(kind: &str, started: Instant) {
+    let waited = started.elapsed().as_secs_f64();
+    if waited > SLOW_WAIT_LOG_SECS {
+        log::info!("GPU lease ({kind}): waited {waited:.1}s for the other model class to drain");
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // One sequential test, not several: the lease is a single global, so
+    // parallel tests interleaving reads and writes on it can hit the very
+    // nested-acquisition deadlock the module comment warns about.
+    #[tokio::test]
+    async fn write_lease_excludes_readers_then_reads_share() {
+        let w = tts_lease().await;
+        // A reader must not acquire while the writer is held.
+        let pending = tokio::spawn(async { drop(llm_lease().await) });
+        tokio::task::yield_now().await;
+        assert!(!pending.is_finished());
+        drop(w);
+        pending.await.expect("reader acquires after writer drops");
+
+        // With no writer queued, read leases are shared.
+        let a = llm_lease().await;
+        let b = llm_lease().await;
+        drop(a);
+        drop(b);
+    }
+}

src/ai/handlers.rs

@@ -468,6 +468,13 @@ pub async fn generate_insight_handler(
         let path_for_task = path.clone();
         let generator_for_task = generator.clone();
         let result = tokio::task::spawn(async move {
+            // Cross-model barrier: if a TTS synthesis holds the GPU, wait it
+            // out BEFORE the generation wall-clock starts. The per-request
+            // lease keeps reqwest budgets honest, but this job-level timeout
+            // would otherwise burn while the first chat call queues behind a
+            // multi-minute synthesis. Dropped immediately — holding it across
+            // the generation would deadlock the chat calls' own leases.
+            drop(crate::ai::gpu::llm_lease().await);
             tokio::time::timeout(
                 std::time::Duration::from_secs(timeout_secs),
                 generator_for_task.generate_insight_for_photo_with_config(
@@ -510,7 +517,9 @@ pub async fn generate_insight_handler(
             }
             Ok(Ok(Err(e))) => {
                 log::error!("Insight generation failed for {}: {:?}", path, e);
-                if let Err(err) = dao.fail_job(&ctx, job_id, &format!("{:?}", e)) {
+                // `{:#}` = one-line context chain; the job's error_message is
+                // returned to the client verbatim, so no Debug/backtrace here.
+                if let Err(err) = dao.fail_job(&ctx, job_id, &format!("{:#}", e)) {
                     log::error!("Failed to mark job {} as failed: {:?}", job_id, err);
                 }
             }
@@ -844,6 +853,9 @@ pub async fn generate_agentic_insight_handler(
         let path_for_task = path.clone();
         let generator_for_task = generator.clone();
         let result = tokio::task::spawn(async move {
+            // Cross-model barrier — see generate_insight_handler: wait out any
+            // running TTS synthesis before the generation wall-clock starts.
+            drop(crate::ai::gpu::llm_lease().await);
             tokio::time::timeout(
                 std::time::Duration::from_secs(timeout_secs),
                 generator_for_task.generate_agentic_insight_for_photo(
@@ -884,7 +896,9 @@ pub async fn generate_agentic_insight_handler(
             }
             Ok(Ok(Err(e))) => {
                 log::error!("Agentic insight generation failed for {}: {:?}", path, e);
-                if let Err(err) = dao.fail_job(&ctx, job_id, &format!("{:?}", e)) {
+                // `{:#}` = one-line context chain; the job's error_message is
+                // returned to the client verbatim, so no Debug/backtrace here.
+                if let Err(err) = dao.fail_job(&ctx, job_id, &format!("{:#}", e)) {
                     log::error!("Failed to mark job {} as failed: {:?}", job_id, err);
                 }
             }

src/ai/insight_generator.rs

@@ -33,30 +33,40 @@ use crate::utils::{earliest_fs_time, normalize_path};
 /// and labels the truncation via `found_header`.
 const LOCATION_HISTORY_DISPLAY_LIMIT: usize = 20;
 
+/// Strip common markdown decoration (bold/italic markers, heading hashes,
+/// backticks, quotes) from both ends of a model-emitted title. Models wrap
+/// the line despite the prompt: `**Title: A Day in the Woods**`,
+/// `## Title: ...`, `"..."`.
+pub(crate) fn strip_title_markdown(s: &str) -> &str {
+    s.trim_matches(|c: char| matches!(c, '*' | '_' | '`' | '#' | '"') || c.is_whitespace())
+}
+
 /// Parse a "Title: ...\n\n<body>" response into (title, body).
 /// Falls back to the first sentence as the title if the model didn't
 /// follow the format.
 pub(crate) fn parse_title_body(raw: &str) -> (String, String) {
     let trimmed = raw.trim();
 
-    // Try "Title: <title>\n\n<body>" or "Title: <title>\n<body>"
-    if let Some(rest) = trimmed
+    // Try "Title: <title>\n<body>", tolerating markdown decoration around
+    // the title line.
+    let (first_line, rest) = match trimmed.find('\n') {
+        Some(pos) => (&trimmed[..pos], trimmed[pos..].trim()),
+        None => (trimmed, ""),
+    };
+    let first_line = strip_title_markdown(first_line);
+    if let Some(t) = first_line
         .strip_prefix("Title:")
-        .or_else(|| trimmed.strip_prefix("title:"))
+        .or_else(|| first_line.strip_prefix("title:"))
     {
-        let rest = rest.trim_start();
-        if let Some(split_pos) = rest.find("\n\n").or_else(|| rest.find('\n')) {
-            let title = rest[..split_pos].trim();
-            let body = rest[split_pos..].trim();
-            if !title.is_empty() && !body.is_empty() {
-                return (title.to_string(), body.to_string());
-            }
+        let title = strip_title_markdown(t);
+        if !title.is_empty() && !rest.is_empty() {
+            return (title.to_string(), rest.to_string());
         }
     }
 
     // Fallback: first sentence (up to first `. ` or `.\n`) becomes the title
     if let Some(pos) = trimmed.find(". ").or_else(|| trimmed.find(".\n")) {
-        let title = &trimmed[..pos];
+        let title = strip_title_markdown(&trimmed[..pos]);
         let body = trimmed[pos + 1..].trim();
         if title.len() <= 100 && !body.is_empty() {
             return (title.to_string(), body.to_string());
@@ -65,7 +75,7 @@ pub(crate) fn parse_title_body(raw: &str) -> (String, String) {
 
     // Last resort: truncate to 60 chars for title, full text as body
     let title: String = trimmed.chars().take(60).collect();
-    let title = title.trim_end().to_string();
+    let title = strip_title_markdown(title.trim_end()).to_string();
     (title, trimmed.to_string())
 }
 
@@ -535,7 +545,7 @@ impl InsightGenerator {
         // (`LLM_BACKEND` switch). Must match the backend that populated the
         // daily-summary embeddings or similarity search will be garbage.
         let query_embedding =
-            crate::ai::embed_one(&self.ollama, self.llamacpp.as_deref(), &query).await?;
+            crate::ai::embed_query(&self.ollama, self.llamacpp.as_deref(), &query).await?;
 
         // Search for similar daily summaries with time-based weighting
         // This prioritizes summaries temporally close to the query date
@@ -575,6 +585,67 @@ impl InsightGenerator {
         Ok(formatted)
     }
 
+    /// Semantic search over daily summaries for the agentic `search_rag`
+    /// tool. Embeds the caller's query as-is (no metadata boilerplate) and
+    /// only applies time weighting when an anchor date is provided —
+    /// without one, results rank purely by similarity across all time.
+    async fn search_summaries_semantic(
+        &self,
+        query: &str,
+        date: Option<chrono::NaiveDate>,
+        limit: usize,
+    ) -> Result<Vec<String>> {
+        let tracer = global_tracer();
+        let current_cx = opentelemetry::Context::current();
+        let mut span = tracer.start_with_context("ai.rag.search_daily_summaries", &current_cx);
+        span.set_attribute(KeyValue::new("query", query.to_string()));
+        span.set_attribute(KeyValue::new("limit", limit as i64));
+        span.set_attribute(KeyValue::new("time_weighted", date.is_some()));
+        if let Some(d) = date {
+            span.set_attribute(KeyValue::new("date", d.to_string()));
+        }
+        let search_cx = current_cx.with_span(span);
+
+        log::info!("RAG QUERY: {} (anchor date: {:?})", query, date);
+
+        // Must use the same backend that populated the daily-summary
+        // embeddings or similarity search is garbage (see embed_one docs).
+        let query_embedding =
+            crate::ai::embed_query(&self.ollama, self.llamacpp.as_deref(), query).await?;
+
+        let mut summary_dao = self
+            .daily_summary_dao
+            .lock()
+            .expect("Unable to lock DailySummaryDao");
+
+        let similar_summaries = match date {
+            Some(d) => summary_dao.find_similar_summaries_with_time_weight(
+                &search_cx,
+                &query_embedding,
+                &d.format("%Y-%m-%d").to_string(),
+                limit,
+            ),
+            None => summary_dao.find_similar_summaries(&search_cx, &query_embedding, limit),
+        }
+        .map_err(|e| anyhow::anyhow!("Failed to find similar summaries: {:?}", e))?;
+
+        search_cx.span().set_attribute(KeyValue::new(
+            "results_count",
+            similar_summaries.len() as i64,
+        ));
+        search_cx.span().set_status(Status::Ok);
+
+        Ok(similar_summaries
+            .into_iter()
+            .map(|s| {
+                format!(
+                    "[{}] {} ({} messages):\n{}",
+                    s.date, s.contact, s.message_count, s.summary
+                )
+            })
+            .collect())
+    }
+
     /// Build a metadata-based query (fallback when no topics available)
     fn build_metadata_query(
         date: chrono::NaiveDate,
@@ -626,7 +697,7 @@ impl InsightGenerator {
         let calendar_cx = parent_cx.with_span(span);
 
         let query_embedding = if let Some(loc) = location {
-            match crate::ai::embed_one(&self.ollama, self.llamacpp.as_deref(), loc).await {
+            match crate::ai::embed_query(&self.ollama, self.llamacpp.as_deref(), loc).await {
                 Ok(emb) => Some(emb),
                 Err(e) => {
                     log::warn!("Failed to generate embedding for location '{}': {}", loc, e);
@@ -798,7 +869,8 @@ impl InsightGenerator {
         };
 
         let query_embedding =
-            match crate::ai::embed_one(&self.ollama, self.llamacpp.as_deref(), &query_text).await {
+            match crate::ai::embed_query(&self.ollama, self.llamacpp.as_deref(), &query_text).await
+            {
                 Ok(emb) => emb,
                 Err(e) => {
                     log::warn!("Failed to generate search embedding: {}", e);
@@ -1737,13 +1809,12 @@ Return ONLY the summary, nothing else."#,
             Some(q) => q.to_string(),
             None => return "Error: missing required parameter 'query'".to_string(),
         };
-        let date_str = match args.get("date").and_then(|v| v.as_str()) {
-            Some(d) => d,
-            None => return "Error: missing required parameter 'date'".to_string(),
-        };
-        let date = match NaiveDate::parse_from_str(date_str, "%Y-%m-%d") {
-            Ok(d) => d,
-            Err(e) => return format!("Error: failed to parse date '{}': {}", date_str, e),
+        let date = match args.get("date").and_then(|v| v.as_str()) {
+            Some(d) => match NaiveDate::parse_from_str(d, "%Y-%m-%d") {
+                Ok(d) => Some(d),
+                Err(e) => return format!("Error: failed to parse date '{}': {}", d, e),
+            },
+            None => None,
         };
         let contact = args
             .get("contact")
@@ -1756,7 +1827,7 @@ Return ONLY the summary, nothing else."#,
             .clamp(1, 25) as usize;
 
         log::info!(
-            "tool_search_rag: query='{}', date={}, contact={:?}, limit={}",
+            "tool_search_rag: query='{}', date={:?}, contact={:?}, limit={}",
             query,
             date,
             contact,
@@ -1777,15 +1848,17 @@ Return ONLY the summary, nothing else."#,
             limit
         };
 
+        // Embed the model's query verbatim — a soft contact bias is the
+        // only decoration. The metadata boilerplate ("On <date>, it was a
+        // <weekday>") that find_relevant_messages_rag prepends drowns the
+        // semantic signal, so the tool path deliberately bypasses it.
+        let search_query = match contact.as_deref() {
+            Some(c) => format!("{} (conversation with {})", query, c),
+            None => query.clone(),
+        };
+
         let results = match self
-            .find_relevant_messages_rag(
-                date,
-                None,
-                contact.as_deref(),
-                None,
-                candidate_limit,
-                Some(&query),
-            )
+            .search_summaries_semantic(&search_query, date, candidate_limit)
             .await
         {
             Ok(results) if !results.is_empty() => results,
@@ -2062,12 +2135,15 @@ Return ONLY the summary, nothing else."#,
     /// Render a list of [`SmsSearchHit`] for the LLM. Prefers the SMS-API
     /// snippet (which already excerpts the matched span and is the only
     /// preview MMS-attachment-only matches have) over the full body, and
-    /// strips the `<mark>` tags the snippet ships with.
+    /// strips the `<mark>` tags the snippet ships with. Each line names
+    /// both parties (`sender → recipient`) — results can span multiple
+    /// conversations, and a sender-only label leaves sent messages
+    /// unattributable to a thread.
     fn format_search_hits(hits: &[SmsSearchHit], mode: &str, date_filtered: bool) -> String {
         let user_name = user_display_name();
         let mut out = String::new();
         out.push_str(&format!(
-            "Found {} messages (mode: {}{}):\n\n",
+            "Found {} messages (mode: {}{}, sender → recipient):\n\n",
             hits.len(),
             mode,
             if date_filtered { ", date-filtered" } else { "" }
@@ -2076,10 +2152,10 @@ Return ONLY the summary, nothing else."#,
             let date = chrono::DateTime::from_timestamp(h.date, 0)
                 .map(|dt| dt.format("%Y-%m-%d").to_string())
                 .unwrap_or_else(|| h.date.to_string());
-            let direction: &str = if h.type_ == 2 {
-                &user_name
+            let direction = if h.type_ == 2 {
+                format!("{} → {}", user_name, h.contact_name)
             } else {
-                &h.contact_name
+                format!("{} → {}", h.contact_name, user_name)
             };
             let score = h
                 .similarity_score
@@ -2150,11 +2226,18 @@ Return ONLY the summary, nothing else."#,
         {
             Ok(messages) if !messages.is_empty() => {
                 let user_name = user_display_name();
+                // Name both parties — without a contact filter the window
+                // spans every conversation, and a sender-only label leaves
+                // sent messages unattributable to a thread.
                 let formatted: Vec<String> = messages
                     .iter()
                     .take(limit)
                     .map(|m| {
-                        let sender: &str = if m.is_sent { &user_name } else { &m.contact };
+                        let direction = if m.is_sent {
+                            format!("{} → {}", user_name, m.contact)
+                        } else {
+                            format!("{} → {}", m.contact, user_name)
+                        };
                         let ts = DateTime::from_timestamp(m.timestamp, 0)
                             .map(|dt| {
                                 dt.with_timezone(&Local)
@@ -2162,7 +2245,7 @@ Return ONLY the summary, nothing else."#,
                                     .to_string()
                             })
                             .unwrap_or_else(|| "unknown".to_string());
-                        format!("[{}] {}: {}", ts, sender, m.body)
+                        format!("[{}] {}: {}", ts, direction, m.body)
                     })
                     .collect();
                 format!(
@@ -2870,17 +2953,34 @@ Return ONLY the summary, nothing else."#,
         // Generate embedding for name + description (best-effort) via the
         // configured local backend.
         let embed_text = format!("{} {}", name, description);
-        let embedding: Option<Vec<u8>> =
-            match crate::ai::embed_one(&self.ollama, self.llamacpp.as_deref(), &embed_text).await {
-                Ok(vec) => {
-                    let bytes: Vec<u8> = vec.iter().flat_map(|f| f.to_le_bytes()).collect();
-                    Some(bytes)
-                }
-                Err(e) => {
-                    log::warn!("Embedding generation failed for entity '{}': {}", name, e);
-                    None
-                }
-            };
+        let embedding: Option<Vec<u8>> = match crate::ai::embed_document(
+            &self.ollama,
+            self.llamacpp.as_deref(),
+            &embed_text,
+        )
+        .await
+        {
+            // The entities table has no dim check at the DAO layer, and a
+            // wrong-dim vector silently kills dedup/recall (cosine over
+            // mismatched lengths is 0) — guard here, store None instead.
+            Ok(vec) if vec.len() == crate::ai::embedding_dim() => {
+                let bytes: Vec<u8> = vec.iter().flat_map(|f| f.to_le_bytes()).collect();
+                Some(bytes)
+            }
+            Ok(vec) => {
+                log::warn!(
+                    "Entity '{}' embedding has {} dims (expected {}) — storing without embedding",
+                    name,
+                    vec.len(),
+                    crate::ai::embedding_dim()
+                );
+                None
+            }
+            Err(e) => {
+                log::warn!("Embedding generation failed for entity '{}': {}", name, e);
+                None
+            }
+        };
 
         let now = chrono::Utc::now().timestamp();
         let insert = InsertEntity {
@@ -3206,21 +3306,25 @@ Return ONLY the summary, nothing else."#,
         if opts.daily_summaries_present {
             tools.push(Tool::function(
                 "search_rag",
-                "Date-anchored semantic search over the user's daily-summary corpus. \
-                 Returns up to `limit` summaries most semantically similar to `query`, \
-                 weighted toward summaries near `date`. For raw message text across all \
-                 time, prefer `search_messages`. \
-                 Examples: `{query: \"family dinner\", date: \"2018-12-24\"}` — what \
+                "Semantic search over the user's daily-summary corpus. Returns up to \
+                 `limit` summaries most semantically similar to `query`. Pass `date` \
+                 to anchor in time: summaries near that date rank higher and matches \
+                 months away decay sharply. Omit `date` to rank purely by semantic \
+                 similarity across all time — do this for \"when did X happen?\" \
+                 questions where the date is unknown. For raw message text, prefer \
+                 `search_messages`. \
+                 Examples: `{query: \"family dinner\"}` — best matches across all \
+                 time. `{query: \"family dinner\", date: \"2018-12-24\"}` — what \
                  daily summaries near Christmas Eve mention family / dinner / gathering. \
                  `{query: \"work travel\", date: \"2019-06-15\", contact: \"Alice\"}` — \
-                 narrowed to summaries that involve Alice.",
+                 biased toward summaries that involve Alice.",
                 serde_json::json!({
                     "type": "object",
-                    "required": ["query", "date"],
+                    "required": ["query"],
                     "properties": {
                         "query": { "type": "string", "description": "Free-text query, semantically matched." },
-                        "date":  { "type": "string", "description": "Anchor date, YYYY-MM-DD. Summaries near this date rank higher." },
-                        "contact": { "type": "string", "description": "Optional contact name to bias toward conversations with that person." },
+                        "date":  { "type": "string", "description": "Optional anchor date, YYYY-MM-DD. When set, summaries near this date rank higher; omit to search all time evenly." },
+                        "contact": { "type": "string", "description": "Optional contact name to bias toward conversations with that person (soft semantic bias, not a hard filter)." },
                         "limit": { "type": "integer", "description": "Max summaries to return (default 10, max 25)." }
                     }
                 }),
@@ -4763,12 +4867,22 @@ mod tests {
         let hit = make_search_hit(1, "Sarah", "see you at the lake tomorrow", None, 1);
         let out = InsightGenerator::format_search_hits(&[hit], "fts5", false);
 
-        assert!(out.starts_with("Found 1 messages (mode: fts5):"));
+        assert!(out.starts_with("Found 1 messages (mode: fts5"));
         assert!(out.contains("see you at the lake tomorrow"));
-        assert!(out.contains("Sarah —"));
+        // Received message: contact is the sender.
+        assert!(out.contains("Sarah →"));
         assert!(!out.contains("date-filtered"));
     }
 
+    #[test]
+    fn format_search_hits_labels_sent_direction() {
+        // Sent messages must name the recipient — results can span multiple
+        // conversations, and a sender-only label left them unattributable.
+        let hit = make_search_hit(5, "Sarah", "on my way", None, 2);
+        let out = InsightGenerator::format_search_hits(&[hit], "fts5", false);
+        assert!(out.contains("→ Sarah —"));
+    }
+
     #[test]
     fn format_search_hits_prefers_snippet_over_body_and_strips_marks() {
         let hit = make_search_hit(
@@ -4799,7 +4913,7 @@ mod tests {
 
         assert!(out.contains("birthday_cake.jpg"));
         assert!(!out.contains("<mark>"));
-        assert!(out.contains("Mom —"));
+        assert!(out.contains("Mom →"));
     }
 
     #[test]
@@ -5022,6 +5136,28 @@ mod tests {
         assert_eq!(b, "Everyone gathered...");
     }
 
+    #[test]
+    fn parse_title_body_strips_bold_wrapper() {
+        let (t, b) = parse_title_body("**Title: A Day in the Woods**\n\nWe hiked the ridge trail.");
+        assert_eq!(t, "A Day in the Woods");
+        assert_eq!(b, "We hiked the ridge trail.");
+    }
+
+    #[test]
+    fn parse_title_body_strips_bold_label_only() {
+        // Bold around just the label: "**Title:** X"
+        let (t, b) = parse_title_body("**Title:** Garden Party\n\nEveryone gathered...");
+        assert_eq!(t, "Garden Party");
+        assert_eq!(b, "Everyone gathered...");
+    }
+
+    #[test]
+    fn parse_title_body_strips_heading_hashes() {
+        let (t, b) = parse_title_body("## Title: Morning Walk\nThe sun was rising...");
+        assert_eq!(t, "Morning Walk");
+        assert_eq!(b, "The sun was rising...");
+    }
+
     #[test]
     fn parse_title_body_fallback_first_sentence() {
         let (t, b) = parse_title_body("A warm summer day. We gathered at the park for a picnic.");

src/ai/llamacpp.rs

@@ -142,6 +142,11 @@ impl LlamaCppClient {
     /// Chatterbox generation knobs are forwarded when set (caller is expected
     /// to have range-clamped them): `exaggeration` (0.25–2.0, emotion),
     /// `cfg_weight` (0.0–1.0, pace), `temperature` (0.05–5.0, randomness).
+    ///
+    /// Callers must hold the GPU write lease (`ai::gpu::tts_lease`) across
+    /// this call. It is taken at the call sites in `ai::tts` rather than here
+    /// so the speech-job path can flip its job to `running` between acquiring
+    /// the GPU and sending the request.
     pub async fn text_to_speech(
         &self,
         input: &str,
@@ -204,6 +209,9 @@ impl LlamaCppClient {
     /// List voices in the Chatterbox voice library (raw JSON passthrough).
     pub async fn list_voices(&self) -> Result<Value> {
         let url = format!("{}/upstream/{}/voices", self.swap_root(), self.tts_model);
+        // The /upstream passthrough spins Chatterbox up (evicting the LLM),
+        // so it takes the exclusive GPU lease like synthesis does.
+        let _gpu = crate::ai::gpu::tts_lease().await;
         let resp = self
             .client
             .get(&url)
@@ -237,6 +245,9 @@ impl LlamaCppClient {
             .text("voice_name", voice_name.to_string())
             .part("voice_file", part);
 
+        // The /upstream passthrough spins Chatterbox up (evicting the LLM),
+        // so it takes the exclusive GPU lease like synthesis does.
+        let _gpu = crate::ai::gpu::tts_lease().await;
         let resp = self
             .client
             .post(&url)
@@ -253,6 +264,37 @@ impl LlamaCppClient {
         resp.json().await.context("parsing create_voice response")
     }
 
+    /// Delete a cloned voice from the Chatterbox voice library
+    /// (`DELETE /voices/{name}` on the upstream, via llama-swap passthrough).
+    pub async fn delete_voice(&self, voice_name: &str) -> Result<Value> {
+        let url = format!(
+            "{}/upstream/{}/voices/{}",
+            self.swap_root(),
+            self.tts_model,
+            voice_name
+        );
+        // The /upstream passthrough spins Chatterbox up (evicting the LLM),
+        // so it takes the exclusive GPU lease like synthesis does.
+        let _gpu = crate::ai::gpu::tts_lease().await;
+        let resp = self
+            .client
+            .delete(&url)
+            .send()
+            .await
+            .with_context(|| format!("DELETE {} failed", url))?;
+
+        if !resp.status().is_success() {
+            let status = resp.status();
+            let text = resp.text().await.unwrap_or_default();
+            bail!("llama-swap delete_voice failed: {} — {}", status, text);
+        }
+        // Some upstreams reply with an empty body on delete.
+        Ok(resp
+            .json()
+            .await
+            .unwrap_or_else(|_| json!({ "status": "deleted" })))
+    }
+
     /// Translate canonical messages to the OpenAI-compatible wire shape.
     /// Behaviorally identical to `OpenRouterClient::messages_to_openai` —
     /// stringify tool-call arguments, rewrite images into content-parts, attach
@@ -453,6 +495,9 @@ impl LlamaCppClient {
             body.insert(k.into(), v);
         }
 
+        // Wait for any TTS synthesis to release the GPU before the request
+        // timeout starts (see ai::gpu).
+        let _gpu = crate::ai::gpu::llm_lease().await;
         let resp = self
             .client
             .post(&url)
@@ -571,6 +616,10 @@ impl LlmClient for LlamaCppClient {
             body.insert(k.into(), v);
         }
 
+        // Wait for any TTS synthesis to release the GPU before the request
+        // timeout starts (see ai::gpu). The guard is moved into the stream
+        // below so the lease spans the whole generation, not just the send.
+        let gpu = crate::ai::gpu::llm_lease().await;
         let resp = self
             .client
             .post(&url)
@@ -587,6 +636,7 @@ impl LlmClient for LlamaCppClient {
 
         let byte_stream = resp.bytes_stream();
         let stream = async_stream::stream! {
+            let _gpu = gpu;
             let mut byte_stream = byte_stream;
             let mut buf: Vec<u8> = Vec::new();
             let mut accumulated_content = String::new();
@@ -702,6 +752,9 @@ impl LlmClient for LlamaCppClient {
     }
 
     async fn generate_embeddings(&self, texts: &[&str]) -> Result<Vec<Vec<f32>>> {
+        // Deliberately NO GPU lease: the embed slot sits in llama-swap's
+        // always-resident group and never participates in a model swap, so
+        // leasing here would only stall searches behind a queued synthesis.
         let url = format!("{}/embeddings", self.base_url);
         let body = json!({
             "model": self.embedding_model,

src/ai/local_llm.rs

@@ -0,0 +1,88 @@
+//! Bundle of the local LLM pair (Ollama + optional llama-swap) with the
+//! `LLM_BACKEND` dispatch baked in.
+//!
+//! Exists because passing the pair around as loose values invited the same
+//! bug three times: import/backfill tooling embedded corpora via
+//! `OllamaClient` directly while the query side dispatched through
+//! `embed_one`, so flipping `LLM_BACKEND=llamacpp` silently split queries
+//! and corpus into different vector spaces. Anything that writes or reads
+//! embeddings should go through this type (or `embed_one`/`embed_many`),
+//! never a concrete client.
+//!
+//! Deliberately knows nothing about chat policy — hybrid/OpenRouter routing
+//! is request-scoped and stays in `ResolvedBackend`. This is only the
+//! local stack: embeddings and offline single-shot generation.
+
+// Constructed by binaries, not the server — dead code from main.rs's view.
+#![allow(dead_code)]
+
+use std::sync::Arc;
+
+use anyhow::Result;
+
+use super::llamacpp::LlamaCppClient;
+use super::llm_client::LlmClient;
+use super::ollama::{EMBEDDING_MODEL, OllamaClient};
+
+#[derive(Clone)]
+pub struct LocalLlm {
+    ollama: OllamaClient,
+    llamacpp: Option<Arc<LlamaCppClient>>,
+}
+
+impl LocalLlm {
+    pub fn new(ollama: OllamaClient, llamacpp: Option<Arc<LlamaCppClient>>) -> Self {
+        Self { ollama, llamacpp }
+    }
+
+    /// Construct from the canonical env wiring shared with `AppState`.
+    pub fn from_env() -> Self {
+        Self::new(
+            crate::state::build_ollama_from_env(),
+            crate::state::build_llamacpp_from_env(),
+        )
+    }
+
+    /// Embed a search query (applies `EMBED_QUERY_PREFIX`). Callers must
+    /// pick query vs document — retrieval models treat the two sides
+    /// differently and an unmarked embed invites prefix-mismatch bugs.
+    pub async fn embed_query(&self, text: &str) -> Result<Vec<f32>> {
+        super::embed_query(&self.ollama, self.llamacpp.as_deref(), text).await
+    }
+
+    /// Embed corpus text (applies `EMBED_DOCUMENT_PREFIX`).
+    pub async fn embed_document(&self, text: &str) -> Result<Vec<f32>> {
+        super::embed_document(&self.ollama, self.llamacpp.as_deref(), text).await
+    }
+
+    /// Single-shot local text generation via the `LLM_BACKEND`-selected
+    /// client (offline tooling; chat turns belong to `ResolvedBackend`).
+    pub async fn generate(&self, prompt: &str, system: Option<&str>) -> Result<String> {
+        if super::local_backend_is_llamacpp() {
+            if let Some(lc) = self.llamacpp.as_deref() {
+                return <LlamaCppClient as LlmClient>::generate(lc, prompt, system, None).await;
+            }
+            anyhow::bail!(
+                "LLM_BACKEND=llamacpp but LlamaCppClient is unconfigured — \
+                 set LLAMA_SWAP_URL or switch to LLM_BACKEND=ollama"
+            );
+        }
+        self.ollama.generate(prompt, system).await
+    }
+
+    /// Label identifying which backend + model produces embeddings right
+    /// now. Store it alongside vectors (`model_version` columns) so a
+    /// backend flip is detectable in the data, not just in env history.
+    pub fn embedding_model_version(&self) -> String {
+        if super::local_backend_is_llamacpp() {
+            let slot = self
+                .llamacpp
+                .as_deref()
+                .map(|c| c.embedding_model.as_str())
+                .unwrap_or("embed");
+            format!("llama-swap:{}", slot)
+        } else {
+            EMBEDDING_MODEL.to_string()
+        }
+    }
+}

src/ai/mod.rs

@@ -3,13 +3,16 @@ pub mod backend;
 pub mod clip_client;
 pub mod daily_summary_job;
 pub mod face_client;
+pub mod gpu;
 pub mod handlers;
 pub mod insight_chat;
 pub mod insight_generator;
 pub mod llamacpp;
 pub mod llm_client;
+pub mod local_llm;
 pub mod ollama;
 pub mod openrouter;
+pub mod pronunciation;
 pub mod sms_client;
 pub mod tts;
 pub mod turn_registry;
@@ -34,11 +37,15 @@ pub use llamacpp::LlamaCppClient;
 pub use llm_client::{
     ChatMessage, LlmClient, ModelCapabilities, Tool, ToolCall, ToolCallFunction, ToolFunction,
 };
+// LocalLlm is constructed by binaries (reembed_embeddings, importers), not the server
+#[allow(unused_imports)]
+pub use local_llm::LocalLlm;
 pub use ollama::{EMBEDDING_MODEL, OllamaClient};
 pub use sms_client::{SmsApiClient, SmsMessage};
 pub use tts::{
-    create_voice_from_library_handler, create_voice_upload_handler, list_voices_handler,
-    tts_speech_handler,
+    cancel_speech_job_handler, create_speech_job_handler, create_voice_from_library_handler,
+    create_voice_upload_handler, delete_voice_handler, list_voices_handler,
+    speech_job_status_handler, tts_speech_handler,
 };
 
 /// Display name used for the user in message transcripts and first-person
@@ -69,35 +76,100 @@ pub fn local_backend_is_llamacpp() -> bool {
     )
 }
 
-/// Embed one string via the configured local backend. Routes through
-/// llama-swap when `LLM_BACKEND=llamacpp` (and a client is configured),
-/// else Ollama. Returns the single embedding vector. See
-/// [`local_backend_is_llamacpp`] for the rationale on consistency.
-pub async fn embed_one(
+/// Expected embedding dimensionality, env-overridable via `EMBEDDING_DIM`
+/// (default 768, nomic-embed-text). Every store/query dim check reads this —
+/// swapping to a different-dim model (e.g. Qwen3-Embedding-0.6B at 1024) is
+/// then a config flip plus a `reembed_embeddings` run, not a code change.
+/// Cached for the process lifetime; a flip requires a restart anyway since
+/// the corpus must be re-embedded with it.
+pub fn embedding_dim() -> usize {
+    static DIM: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
+    *DIM.get_or_init(|| {
+        std::env::var("EMBEDDING_DIM")
+            .ok()
+            .and_then(|v| v.parse().ok())
+            .unwrap_or(768)
+    })
+}
+
+/// Read an embedding prefix from the environment. `.env` values can't hold
+/// real newlines, so a literal `\n` in the value is expanded — Qwen3-style
+/// query instructions need one ("Instruct: ...\nQuery: ").
+fn embed_prefix(key: &str) -> String {
+    std::env::var(key)
+        .map(|v| v.replace("\\n", "\n"))
+        .unwrap_or_default()
+}
+
+/// Embed a search query. Applies `EMBED_QUERY_PREFIX` (default empty) —
+/// retrieval models distinguish query-side from document-side text:
+/// nomic v1.5 wants `search_query: `, Qwen3-Embedding wants
+/// `Instruct: <task>\nQuery: `. Must pair with the document prefix the
+/// corpus was embedded with or similarity degrades.
+pub async fn embed_query(
     ollama: &OllamaClient,
     llamacpp: Option<&LlamaCppClient>,
     text: &str,
 ) -> anyhow::Result<Vec<f32>> {
+    let prefixed = format!("{}{}", embed_prefix("EMBED_QUERY_PREFIX"), text);
+    embed_one(ollama, llamacpp, &prefixed).await
+}
+
+/// Embed corpus text (the stored side of retrieval). Applies
+/// `EMBED_DOCUMENT_PREFIX` (default empty; nomic v1.5 wants
+/// `search_document: `, Qwen3-Embedding wants none).
+pub async fn embed_document(
+    ollama: &OllamaClient,
+    llamacpp: Option<&LlamaCppClient>,
+    text: &str,
+) -> anyhow::Result<Vec<f32>> {
+    let prefixed = format!("{}{}", embed_prefix("EMBED_DOCUMENT_PREFIX"), text);
+    embed_one(ollama, llamacpp, &prefixed).await
+}
+
+/// Embed a batch of strings via the configured local backend. Routes
+/// through llama-swap when `LLM_BACKEND=llamacpp` (and a client is
+/// configured), else Ollama. See [`local_backend_is_llamacpp`] for the
+/// rationale on consistency.
+pub async fn embed_many(
+    ollama: &OllamaClient,
+    llamacpp: Option<&LlamaCppClient>,
+    texts: &[&str],
+) -> anyhow::Result<Vec<Vec<f32>>> {
     if local_backend_is_llamacpp() {
         if let Some(lc) = llamacpp {
-            let mut vecs = <LlamaCppClient as LlmClient>::generate_embeddings(lc, &[text]).await?;
-            return vecs
-                .pop()
-                .ok_or_else(|| anyhow::anyhow!("llama-swap returned no embeddings"));
+            return <LlamaCppClient as LlmClient>::generate_embeddings(lc, texts).await;
         }
         anyhow::bail!(
             "LLM_BACKEND=llamacpp but LlamaCppClient is unconfigured — \
              set LLAMA_SWAP_URL or switch to LLM_BACKEND=ollama"
         );
     }
-    ollama.generate_embedding(text).await
+    ollama.generate_embeddings(texts).await
+}
+
+/// Embed one string via the configured local backend. Single-text
+/// convenience over [`embed_many`].
+pub async fn embed_one(
+    ollama: &OllamaClient,
+    llamacpp: Option<&LlamaCppClient>,
+    text: &str,
+) -> anyhow::Result<Vec<f32>> {
+    let mut vecs = embed_many(ollama, llamacpp, &[text]).await?;
+    vecs.pop()
+        .ok_or_else(|| anyhow::anyhow!("embedding backend returned no embeddings"))
 }
 
 #[cfg(test)]
 mod env_dispatch_tests {
     use super::*;
 
+    /// Env vars are process-global, and the test harness runs in parallel —
+    /// without this lock the `LLM_BACKEND` tests race each other and flake.
+    static ENV_LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
+
     fn with_env<F: FnOnce()>(key: &str, val: Option<&str>, f: F) {
+        let _guard = ENV_LOCK.lock().unwrap_or_else(|p| p.into_inner());
         let prev = std::env::var(key).ok();
         match val {
             Some(v) => unsafe { std::env::set_var(key, v) },

src/ai/ollama.rs

@@ -548,7 +548,16 @@ Capture the key moment or theme. Return ONLY the title, nothing else."#,
         let title = self
             .generate_with_images(&prompt, Some(system), None)
             .await?;
-        Ok(title.trim().trim_matches('"').to_string())
+        // Models decorate despite "Return ONLY the title": quotes, bold
+        // markers, sometimes a "Title:" label.
+        use crate::ai::insight_generator::strip_title_markdown;
+        let cleaned = strip_title_markdown(title.trim());
+        let cleaned = cleaned
+            .strip_prefix("Title:")
+            .or_else(|| cleaned.strip_prefix("title:"))
+            .map(strip_title_markdown)
+            .unwrap_or(cleaned);
+        Ok(cleaned.to_string())
     }
 
     /// Generate a summary for a single photo based on its context
@@ -1046,13 +1055,14 @@ Analyze the image and use specific details from both the visual content and the
             }
         };
 
-        // Validate embedding dimensions (should be 768 for nomic-embed-text:v1.5)
+        // Validate embedding dimensions (EMBEDDING_DIM; 768 for nomic-embed-text:v1.5)
         for (i, embedding) in embeddings.iter().enumerate() {
-            if embedding.len() != 768 {
+            if embedding.len() != crate::ai::embedding_dim() {
                 log::warn!(
-                    "Unexpected embedding dimensions for item {}: {} (expected 768)",
+                    "Unexpected embedding dimensions for item {}: {} (expected {})",
                     i,
-                    embedding.len()
+                    embedding.len(),
+                    crate::ai::embedding_dim()
                 );
             }
         }

src/ai/pronunciation.rs

@@ -0,0 +1,282 @@
+// User-configurable pronunciation overrides for TTS. Chatterbox mispronounces
+// place names ("Worcester"), initialisms ("WSL"), and clipped abbreviations
+// ("blvd"), so we rewrite them to phonetic spellings before synthesis.
+//
+// The map lives in a JSON file on the server — a flat object of
+// `"written form": "spoken form"` pairs, e.g.:
+//
+//   {
+//     "Worcester": "Wuster",
+//     "WSL": "W S L",
+//     "blvd": "boulevard",
+//     "Dr.": "Doctor"
+//   }
+//
+// Path comes from `TTS_PRONUNCIATIONS_PATH` (default `tts_pronunciations.json`
+// in the working directory). A missing file simply disables the feature. The
+// file is re-read whenever its mtime changes, so edits apply to the next
+// synthesis without a restart; a malformed edit keeps the last good map and
+// logs the parse error instead of silently dropping all overrides.
+//
+// Matching rules:
+// - Whole words only — `cat` never rewrites `category`. (Boundaries are only
+//   asserted next to word characters, so keys like `Dr.` still work.)
+// - Smartcase: an all-lowercase key matches case-insensitively; a key with
+//   any uppercase matches exactly. That lets `worcester` catch every casing
+//   while `US` (the country) leaves the pronoun `us` alone.
+// - Longer keys win over shorter ones (`New York Times` before `New York`).
+
+use regex::Regex;
+use std::collections::HashMap;
+use std::path::Path;
+use std::sync::{Arc, LazyLock, Mutex as StdMutex};
+use std::time::SystemTime;
+
+/// A compiled pronunciation map: one alternation regex over every key plus
+/// the lookup tables the replacement closure resolves matches against.
+#[derive(Default)]
+struct CompiledMap {
+    /// `None` when the map is empty — apply() is then a no-op.
+    regex: Option<Regex>,
+    /// Case-sensitive entries, keyed verbatim.
+    exact: HashMap<String, String>,
+    /// Case-insensitive entries, keyed lowercased.
+    folded: HashMap<String, String>,
+}
+
+impl CompiledMap {
+    fn from_entries(entries: &HashMap<String, String>) -> Self {
+        let mut keys: Vec<&str> = entries
+            .keys()
+            .map(|k| k.as_str())
+            .filter(|k| !k.trim().is_empty())
+            .collect();
+        if keys.is_empty() {
+            return Self::default();
+        }
+        // Longest key first so overlapping entries prefer the more specific
+        // one (regex alternation is first-match-wins, not longest-match).
+        keys.sort_by(|a, b| b.len().cmp(&a.len()).then(a.cmp(b)));
+
+        let mut exact = HashMap::new();
+        let mut folded = HashMap::new();
+        let alternatives: Vec<String> = keys
+            .iter()
+            .map(|key| {
+                let escaped = regex::escape(key);
+                // Only assert a word boundary where the key edge is a word
+                // character — `\b` adjacent to punctuation (e.g. the dot in
+                // `Dr.`) would otherwise never match.
+                let lead = if key
+                    .chars()
+                    .next()
+                    .is_some_and(|c| c.is_alphanumeric() || c == '_')
+                {
+                    r"\b"
+                } else {
+                    ""
+                };
+                let trail = if key
+                    .chars()
+                    .last()
+                    .is_some_and(|c| c.is_alphanumeric() || c == '_')
+                {
+                    r"\b"
+                } else {
+                    ""
+                };
+                let case_sensitive = key.chars().any(|c| c.is_uppercase());
+                if case_sensitive {
+                    exact.insert(key.to_string(), entries[*key].clone());
+                    format!("{lead}{escaped}{trail}")
+                } else {
+                    folded.insert(key.to_lowercase(), entries[*key].clone());
+                    format!("{lead}(?i:{escaped}){trail}")
+                }
+            })
+            .collect();
+
+        // Escaped fixed strings can't produce an invalid pattern; if one ever
+        // does, treat the whole map as empty rather than panicking a handler.
+        let pattern = alternatives.join("|");
+        let regex = match Regex::new(&pattern) {
+            Ok(r) => Some(r),
+            Err(e) => {
+                log::error!("pronunciation map failed to compile: {e}");
+                None
+            }
+        };
+        Self {
+            regex,
+            exact,
+            folded,
+        }
+    }
+
+    fn apply(&self, text: &str) -> String {
+        let Some(re) = &self.regex else {
+            return text.to_string();
+        };
+        re.replace_all(text, |caps: &regex::Captures| {
+            let m = &caps[0];
+            self.exact
+                .get(m)
+                .or_else(|| self.folded.get(&m.to_lowercase()))
+                .cloned()
+                // Unreachable in practice — every alternative came from one
+                // of the two maps — but never drop the user's text.
+                .unwrap_or_else(|| m.to_string())
+        })
+        .into_owned()
+    }
+}
+
+struct CacheEntry {
+    mtime: Option<SystemTime>,
+    compiled: Arc<CompiledMap>,
+}
+
+static CACHE: LazyLock<StdMutex<Option<CacheEntry>>> = LazyLock::new(|| StdMutex::new(None));
+
+fn config_path() -> String {
+    std::env::var("TTS_PRONUNCIATIONS_PATH")
+        .ok()
+        .map(|s| s.trim().to_string())
+        .filter(|s| !s.is_empty())
+        .unwrap_or_else(|| "tts_pronunciations.json".to_string())
+}
+
+/// Load the compiled map, re-reading the file only when its mtime changed
+/// since the last call (or it appeared/disappeared). Synthesis is serialized
+/// on a single GPU permit, so a stat per call is noise.
+fn current_map() -> Arc<CompiledMap> {
+    let path_s = config_path();
+    let path = Path::new(&path_s);
+    let mtime = std::fs::metadata(path).and_then(|m| m.modified()).ok();
+
+    let mut cache = CACHE.lock().unwrap();
+    if let Some(entry) = cache.as_ref()
+        && entry.mtime == mtime
+    {
+        return entry.compiled.clone();
+    }
+
+    let compiled = match mtime {
+        None => Arc::new(CompiledMap::default()), // no file → no overrides
+        Some(_) => match std::fs::read_to_string(path)
+            .map_err(anyhow::Error::from)
+            .and_then(|s| Ok(serde_json::from_str::<HashMap<String, String>>(&s)?))
+        {
+            Ok(entries) => {
+                log::info!(
+                    "loaded {} pronunciation override(s) from {path_s}",
+                    entries.len()
+                );
+                Arc::new(CompiledMap::from_entries(&entries))
+            }
+            Err(e) => {
+                log::error!("failed to load pronunciation map {path_s}: {e}");
+                // Keep serving the previous map rather than regressing to
+                // none mid-edit; still record the new mtime so the error
+                // logs once per bad save, not once per synthesis.
+                cache
+                    .as_ref()
+                    .map(|c| c.compiled.clone())
+                    .unwrap_or_default()
+            }
+        },
+    };
+    *cache = Some(CacheEntry {
+        mtime,
+        compiled: compiled.clone(),
+    });
+    compiled
+}
+
+/// Rewrite configured words/abbreviations to their phonetic spellings.
+/// Call on cleaned (post-markdown-strip) text, right before synthesis.
+pub fn apply_pronunciations(text: &str) -> String {
+    current_map().apply(text)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn compile(pairs: &[(&str, &str)]) -> CompiledMap {
+        let entries = pairs
+            .iter()
+            .map(|(k, v)| (k.to_string(), v.to_string()))
+            .collect();
+        CompiledMap::from_entries(&entries)
+    }
+
+    #[test]
+    fn empty_map_is_a_noop() {
+        let m = compile(&[]);
+        assert_eq!(m.apply("nothing changes"), "nothing changes");
+    }
+
+    #[test]
+    fn replaces_whole_words_only() {
+        let m = compile(&[("cat", "kitty")]);
+        assert_eq!(m.apply("the cat sat"), "the kitty sat");
+        // No substring rewrites.
+        assert_eq!(m.apply("the category"), "the category");
+        assert_eq!(m.apply("concatenate"), "concatenate");
+    }
+
+    #[test]
+    fn lowercase_keys_match_any_casing() {
+        let m = compile(&[("worcester", "Wuster")]);
+        assert_eq!(m.apply("Worcester is nice"), "Wuster is nice");
+        assert_eq!(m.apply("in WORCESTER today"), "in Wuster today");
+        assert_eq!(m.apply("worcester sauce"), "Wuster sauce");
+    }
+
+    #[test]
+    fn uppercase_keys_match_case_sensitively() {
+        let m = compile(&[("US", "U S")]);
+        assert_eq!(m.apply("the US economy"), "the U S economy");
+        // The pronoun survives.
+        assert_eq!(m.apply("join us today"), "join us today");
+    }
+
+    #[test]
+    fn keys_with_punctuation_work() {
+        // `\b` is only asserted next to word characters, so the trailing dot
+        // doesn't break matching.
+        let m = compile(&[("Dr.", "Doctor"), ("blvd", "boulevard")]);
+        assert_eq!(
+            m.apply("Dr. Smith on Sunset blvd"),
+            "Doctor Smith on Sunset boulevard"
+        );
+    }
+
+    #[test]
+    fn longer_keys_win_over_shorter() {
+        let m = compile(&[("new york", "Noo York"), ("new york times", "the Times")]);
+        assert_eq!(m.apply("read the new york times"), "read the the Times");
+        assert_eq!(m.apply("visit new york soon"), "visit Noo York soon");
+    }
+
+    #[test]
+    fn multiple_occurrences_all_rewrite() {
+        let m = compile(&[("wsl", "W S L")]);
+        assert_eq!(m.apply("WSL and wsl and Wsl"), "W S L and W S L and W S L");
+    }
+
+    #[test]
+    fn replacement_text_is_verbatim() {
+        // Replacements aren't re-scanned — a value containing another key
+        // doesn't cascade.
+        let m = compile(&[("a1", "b2"), ("b2", "c3")]);
+        assert_eq!(m.apply("a1"), "b2");
+    }
+
+    #[test]
+    fn blank_keys_are_ignored() {
+        let m = compile(&[("", "x"), ("  ", "y"), ("ok", "fine")]);
+        assert_eq!(m.apply("ok then"), "fine then");
+    }
+}

src/ai/tts.rs

@@ -6,7 +6,7 @@
 // (audio read directly; video has its audio track extracted via ffmpeg).
 
 use actix_multipart::Multipart;
-use actix_web::{HttpRequest, HttpResponse, Responder, get, post, web};
+use actix_web::{HttpRequest, HttpResponse, Responder, delete, get, post, web};
 use anyhow::Context;
 use base64::Engine;
 use bytes::{BufMut, BytesMut};
@@ -15,10 +15,13 @@ use opentelemetry::KeyValue;
 use opentelemetry::trace::{Span, Status, Tracer};
 use regex::Regex;
 use serde::{Deserialize, Serialize};
-use serde_json::json;
+use serde_json::{Value, json};
+use std::collections::HashMap;
 use std::path::Path;
-use std::sync::LazyLock;
+use std::sync::{LazyLock, Mutex as StdMutex};
+use std::time::{Duration, Instant};
 use tokio::sync::Semaphore;
+use uuid::Uuid;
 
 use crate::data::Claims;
 use crate::file_types::{is_audio_file, is_video_file};
@@ -40,6 +43,105 @@ const MAX_VOICE_UPLOAD_BYTES: usize = 25 * 1024 * 1024; // 25 MB
 /// finishes — that's a wrapper limitation; the chunked-queue plan fixes it.)
 static TTS_PERMIT: LazyLock<Semaphore> = LazyLock::new(|| Semaphore::new(1));
 
+// --- Voice-list cache --------------------------------------------------------
+
+/// Cached raw voice-library JSON. llama-swap's `/upstream/<model>/voices`
+/// passthrough spins the TTS model up just to answer a listing — which can
+/// evict the resident LLM — so we serve a cached copy and only hit upstream on
+/// a cold cache, an explicit `?refresh=1`, or after a voice create/delete
+/// invalidates it (the TTS model is already loaded right then anyway).
+static VOICES_CACHE: LazyLock<StdMutex<Option<Value>>> = LazyLock::new(|| StdMutex::new(None));
+
+fn cached_voices() -> Option<Value> {
+    VOICES_CACHE.lock().unwrap().clone()
+}
+
+fn store_voices_cache(v: &Value) {
+    *VOICES_CACHE.lock().unwrap() = Some(v.clone());
+}
+
+fn invalidate_voices_cache() {
+    *VOICES_CACHE.lock().unwrap() = None;
+}
+
+// --- Async speech jobs -------------------------------------------------------
+//
+// Synthesizing a long insight can take minutes — too long to hang one HTTP
+// request from a phone that may background the app or drop the connection.
+// Durable variant: POST /tts/speech/jobs returns a job id immediately, the
+// synth runs in a spawned task (queuing on TTS_PERMIT instead of fast-failing
+// 429), and the client polls GET /tts/speech/jobs/{id} until it collects the
+// audio. State is in-memory only (deliberately lighter than the chat
+// TurnRegistry): a restart loses jobs, the client surfaces that and retries.
+
+#[derive(Clone, Copy, PartialEq, Eq, Debug, Serialize)]
+#[serde(rename_all = "snake_case")]
+pub enum TtsJobStatus {
+    Queued,
+    Running,
+    Done,
+    Error,
+    Cancelled,
+}
+
+impl TtsJobStatus {
+    fn is_terminal(self) -> bool {
+        matches!(self, Self::Done | Self::Error | Self::Cancelled)
+    }
+}
+
+struct TtsJob {
+    status: TtsJobStatus,
+    format: String,
+    audio_base64: Option<String>,
+    error: Option<String>,
+    created_at: Instant,
+    finished_at: Option<Instant>,
+    abort: Option<tokio::task::AbortHandle>,
+}
+
+/// Finished jobs linger so a client that lost connectivity can still collect
+/// the result on a later poll; anything older than MAX_AGE is dropped outright
+/// (aborted first if somehow still running). Swept lazily on each dispatch.
+const TTS_JOB_RESULT_TTL: Duration = Duration::from_secs(10 * 60);
+const TTS_JOB_MAX_AGE: Duration = Duration::from_secs(30 * 60);
+
+static TTS_JOBS: LazyLock<StdMutex<HashMap<Uuid, TtsJob>>> =
+    LazyLock::new(|| StdMutex::new(HashMap::new()));
+
+fn sweep_stale_jobs(jobs: &mut HashMap<Uuid, TtsJob>, now: Instant) {
+    jobs.retain(|_, job| {
+        let result_expired = job
+            .finished_at
+            .is_some_and(|t| now.duration_since(t) >= TTS_JOB_RESULT_TTL);
+        let too_old = now.duration_since(job.created_at) >= TTS_JOB_MAX_AGE;
+        if too_old && let Some(h) = job.abort.take() {
+            h.abort();
+        }
+        !(result_expired || too_old)
+    });
+}
+
+/// Run `f` against a job, if it still exists.
+fn with_job<R>(id: Uuid, f: impl FnOnce(&mut TtsJob) -> R) -> Option<R> {
+    TTS_JOBS.lock().unwrap().get_mut(&id).map(f)
+}
+
+/// Move a job to a terminal state (first terminal write wins — a cancel that
+/// raced a completion keeps the cancel).
+fn finish_job(id: Uuid, status: TtsJobStatus, audio_base64: Option<String>, error: Option<String>) {
+    with_job(id, |job| {
+        if job.status.is_terminal() {
+            return;
+        }
+        job.status = status;
+        job.audio_base64 = audio_base64;
+        job.error = error;
+        job.finished_at = Some(Instant::now());
+        job.abort = None;
+    });
+}
+
 /// Sanitize a user-supplied voice name. The name is forwarded to Chatterbox
 /// where it becomes a filename in the voice-library directory, so we restrict
 /// it to a safe charset (alphanumerics, dash, underscore) — no path
@@ -64,6 +166,66 @@ fn sanitize_voice_name(raw: &str) -> Option<String> {
     Some(cleaned.chars().take(64).collect())
 }
 
+/// Reference-clip cap in seconds for voice cloning. Chatterbox is zero-shot —
+/// a clean ~10–20s sample is the sweet spot and more rarely helps. Tune via
+/// `LLAMA_SWAP_TTS_REF_SECONDS` (default 30).
+fn tts_ref_seconds() -> u32 {
+    std::env::var("LLAMA_SWAP_TTS_REF_SECONDS")
+        .ok()
+        .and_then(|s| s.trim().parse::<u32>().ok())
+        .filter(|n| *n > 0)
+        .unwrap_or(30)
+}
+
+/// Tag a (sanitized) voice name with the reference window used to create it:
+/// `grandma` → `grandma-30s` (from the start), or `grandma-at1m32s-30s` (30s
+/// window starting at 1:32). The tag makes the window visible in the voice
+/// list so clones of the same source from different sections can be compared.
+/// Skips the append when the name already ends in the same tag; keeps the
+/// 64-char bound by truncating the base name, never the tag.
+fn append_ref_window(name: &str, start: f64, secs: u32) -> String {
+    let start_whole = start.round().max(0.0) as u64;
+    let suffix = if start_whole > 0 {
+        // ':' isn't in the safe voice-name charset, so 1:32 becomes 1m32s.
+        let at = if start_whole >= 60 {
+            format!("at{}m{:02}s", start_whole / 60, start_whole % 60)
+        } else {
+            format!("at{start_whole}s")
+        };
+        format!("-{at}-{secs}s")
+    } else {
+        format!("-{secs}s")
+    };
+    if name.ends_with(&suffix) {
+        return name.to_string();
+    }
+    let max_base = 64usize.saturating_sub(suffix.len());
+    let base: String = name.chars().take(max_base).collect();
+    let base = base.trim_end_matches('-');
+    format!("{base}{suffix}")
+}
+
+/// Resolve a caller-supplied reference window into concrete `(start, duration)`
+/// seconds for ffmpeg. Start defaults to 0; duration defaults to the
+/// `tts_ref_seconds` cap and is clamped to it (the cap is the most audio the
+/// TTS backend benefits from, so longer requests are quietly bounded rather
+/// than rejected). Non-finite or negative values are the caller's bug → Err.
+fn resolve_ref_window(
+    start_seconds: Option<f64>,
+    duration_seconds: Option<f64>,
+) -> Result<(f64, f64), String> {
+    let cap = f64::from(tts_ref_seconds());
+    let start = start_seconds.unwrap_or(0.0);
+    if !start.is_finite() || start < 0.0 {
+        return Err("start_seconds must be a non-negative number".to_string());
+    }
+    let duration = duration_seconds.unwrap_or(cap);
+    if !duration.is_finite() || duration <= 0.0 {
+        return Err("duration_seconds must be a positive number".to_string());
+    }
+    Ok((start, duration.min(cap)))
+}
+
 /// Optional default voice for synthesis when the request doesn't name one.
 /// Set `LLAMA_SWAP_TTS_VOICE=m` to read insights in a cloned voice by default.
 fn default_voice() -> Option<String> {
@@ -125,33 +287,42 @@ fn clean_for_tts(input: &str) -> String {
     s.trim().to_string()
 }
 
+/// Full text-preparation pipeline for synthesis: markdown/emoji cleanup, then
+/// the user's pronunciation overrides (see [`crate::ai::pronunciation`]) on
+/// the resulting plain text — after cleanup so word boundaries aren't
+/// obscured by `**WSL**`-style markup.
+fn prepare_for_tts(input: &str) -> String {
+    crate::ai::pronunciation::apply_pronunciations(&clean_for_tts(input))
+}
+
 /// Decode an audio/video file to mono 24 kHz WAV via ffmpeg, returning the WAV
 /// bytes. Chatterbox validates the reference clip by file *extension* and
 /// rejects several formats (e.g. `.aac`, `.opus`), so we always normalize to
-/// WAV regardless of the source container. Capped at 30s — references only need
-/// a few seconds of clean speech.
-async fn run_ffmpeg_to_wav(input_path: &str) -> anyhow::Result<Vec<u8>> {
+/// WAV regardless of the source container. Extracts `duration` seconds starting
+/// at `start` (see resolve_ref_window) — references only need a few seconds of
+/// clean speech, which may sit anywhere in a long recording.
+async fn run_ffmpeg_to_wav(input_path: &str, start: f64, duration: f64) -> anyhow::Result<Vec<u8>> {
     let out = tempfile::Builder::new()
         .suffix(".wav")
         .tempfile()
         .context("creating temp wav")?;
     let out_s = out.path().to_string_lossy().to_string();
 
-    // Cap the reference clip length. Chatterbox is zero-shot — a clean ~10–20s
-    // sample is the sweet spot and more rarely helps — so we use the first N
-    // seconds. Tune via LLAMA_SWAP_TTS_REF_SECONDS (default 30).
-    let secs = std::env::var("LLAMA_SWAP_TTS_REF_SECONDS")
-        .ok()
-        .and_then(|s| s.trim().parse::<u32>().ok())
-        .filter(|n| *n > 0)
-        .unwrap_or(30)
-        .to_string();
+    let start_s = format!("{start}");
+    let secs = format!("{duration}");
+
+    // -ss before -i is input seeking: fast, and frame accuracy doesn't matter
+    // for picking a speech window.
+    let mut args: Vec<&str> = vec!["-y"];
+    if start > 0.0 {
+        args.extend(["-ss", &start_s]);
+    }
+    args.extend([
+        "-i", input_path, "-vn", "-ac", "1", "-ar", "24000", "-t", &secs, "-f", "wav", &out_s,
+    ]);
 
     let output = tokio::process::Command::new("ffmpeg")
-        .args([
-            "-y", "-i", input_path, "-vn", "-ac", "1", "-ar", "24000", "-t", &secs, "-f", "wav",
-            &out_s,
-        ])
+        .args(&args)
         .output()
         .await
         .context("spawning ffmpeg")?;
@@ -164,7 +335,12 @@ async fn run_ffmpeg_to_wav(input_path: &str) -> anyhow::Result<Vec<u8>> {
 
 /// Normalize in-memory upload bytes to WAV: write to a temp file (keeping the
 /// source extension as an ffmpeg probe hint) then transcode.
-async fn transcode_bytes_to_wav(input: &[u8], src_ext: Option<&str>) -> anyhow::Result<Vec<u8>> {
+async fn transcode_bytes_to_wav(
+    input: &[u8],
+    src_ext: Option<&str>,
+    start: f64,
+    duration: f64,
+) -> anyhow::Result<Vec<u8>> {
     let suffix = src_ext
         .filter(|e| !e.is_empty())
         .map(|e| format!(".{e}"))
@@ -174,7 +350,7 @@ async fn transcode_bytes_to_wav(input: &[u8], src_ext: Option<&str>) -> anyhow::
         .tempfile()
         .context("creating temp input")?;
     std::fs::write(in_tmp.path(), input).context("writing temp input")?;
-    run_ffmpeg_to_wav(&in_tmp.path().to_string_lossy()).await
+    run_ffmpeg_to_wav(&in_tmp.path().to_string_lossy(), start, duration).await
 }
 
 #[derive(Debug, Deserialize)]
@@ -214,7 +390,7 @@ pub async fn tts_speech_handler(
     let parent_context = extract_context_from_request(&http_request);
     let mut span = global_tracer().start_with_context("http.tts.speech", &parent_context);
 
-    let text = clean_for_tts(&req.text);
+    let text = prepare_for_tts(&req.text);
     if text.is_empty() {
         span.set_status(Status::error("text is required"));
         return HttpResponse::BadRequest().json(json!({ "error": "text is required" }));
@@ -255,6 +431,10 @@ pub async fn tts_speech_handler(
         }));
     };
 
+    // Wait for the LLM side to release the GPU before sending — the synthesis
+    // timeout starts at send, not here (see ai::gpu).
+    let _gpu = crate::ai::gpu::tts_lease().await;
+
     match client
         .text_to_speech(&text, voice, format, exaggeration, cfg_weight, temperature)
         .await
@@ -276,16 +456,283 @@ pub async fn tts_speech_handler(
     }
 }
 
-/// GET /tts/voices — list the Chatterbox voice library (raw passthrough).
+#[derive(Debug, Serialize)]
+pub struct TtsJobCreatedResponse {
+    pub job_id: String,
+    pub status: TtsJobStatus,
+}
+
+#[derive(Debug, Serialize)]
+pub struct TtsJobStatusResponse {
+    pub job_id: String,
+    pub status: TtsJobStatus,
+    pub format: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub audio_base64: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub error: Option<String>,
+}
+
+/// POST /tts/speech/jobs — durable variant of /tts/speech for long syntheses.
+/// Returns 202 + a job id immediately; the synth queues on the single GPU
+/// permit (instead of fast-failing 429) and the client polls the job until
+/// the audio is ready.
+#[post("/tts/speech/jobs")]
+pub async fn create_speech_job_handler(
+    http_request: HttpRequest,
+    _claims: Claims,
+    req: web::Json<TtsSpeechRequest>,
+    app_state: web::Data<AppState>,
+) -> impl Responder {
+    let parent_context = extract_context_from_request(&http_request);
+    let mut span =
+        global_tracer().start_with_context("http.tts.speech_job.create", &parent_context);
+
+    let text = prepare_for_tts(&req.text);
+    if text.is_empty() {
+        span.set_status(Status::error("text is required"));
+        return HttpResponse::BadRequest().json(json!({ "error": "text is required" }));
+    }
+    if app_state.llamacpp.is_none() {
+        span.set_status(Status::error("tts backend not configured"));
+        return HttpResponse::ServiceUnavailable()
+            .json(json!({ "error": "TTS backend not configured (set LLAMA_SWAP_URL)" }));
+    }
+
+    let format = req
+        .format
+        .as_deref()
+        .filter(|s| !s.is_empty())
+        .unwrap_or("mp3")
+        .to_string();
+    let voice = req
+        .voice
+        .clone()
+        .filter(|s| !s.is_empty())
+        .or_else(default_voice);
+    // Clamp generation knobs to Chatterbox's documented ranges before forwarding.
+    let exaggeration = req.exaggeration.map(|x| x.clamp(0.25, 2.0));
+    let cfg_weight = req.cfg_weight.map(|x| x.clamp(0.0, 1.0));
+    let temperature = req.temperature.map(|x| x.clamp(0.05, 5.0));
+
+    span.set_attribute(KeyValue::new("tts.format", format.clone()));
+    span.set_attribute(KeyValue::new("tts.has_voice", voice.is_some()));
+    span.set_attribute(KeyValue::new("tts.text_len", text.len() as i64));
+
+    let job_id = Uuid::new_v4();
+    {
+        let mut jobs = TTS_JOBS.lock().unwrap();
+        sweep_stale_jobs(&mut jobs, Instant::now());
+        jobs.insert(
+            job_id,
+            TtsJob {
+                status: TtsJobStatus::Queued,
+                format: format.clone(),
+                audio_base64: None,
+                error: None,
+                created_at: Instant::now(),
+                finished_at: None,
+                abort: None,
+            },
+        );
+    }
+
+    let state = app_state.clone();
+    let handle = tokio::spawn(async move {
+        // Queue rather than fast-fail: jobs wait their turn for the GPU.
+        let _permit = match TTS_PERMIT.acquire().await {
+            Ok(p) => p,
+            Err(_) => {
+                finish_job(
+                    job_id,
+                    TtsJobStatus::Error,
+                    None,
+                    Some("TTS queue closed".to_string()),
+                );
+                return;
+            }
+        };
+        // Wait for the LLM side to release the GPU too (see ai::gpu) — only
+        // then does the job count as running. The synthesis timeout starts at
+        // the HTTP send below, so neither wait burns it, and the client can
+        // anchor its own deadline to the queued→running transition.
+        let _gpu = crate::ai::gpu::tts_lease().await;
+
+        // Cancelled while queued — release the permits without synthesizing.
+        let cancelled = with_job(job_id, |job| {
+            if job.status == TtsJobStatus::Queued {
+                job.status = TtsJobStatus::Running;
+                false
+            } else {
+                true
+            }
+        })
+        .unwrap_or(true);
+        if cancelled {
+            return;
+        }
+
+        let Some(client) = state.llamacpp.as_ref() else {
+            finish_job(
+                job_id,
+                TtsJobStatus::Error,
+                None,
+                Some("TTS backend not configured".to_string()),
+            );
+            return;
+        };
+        match client
+            .text_to_speech(
+                &text,
+                voice.as_deref(),
+                &format,
+                exaggeration,
+                cfg_weight,
+                temperature,
+            )
+            .await
+        {
+            Ok(bytes) => {
+                let audio = base64::engine::general_purpose::STANDARD.encode(&bytes);
+                finish_job(job_id, TtsJobStatus::Done, Some(audio), None);
+            }
+            Err(e) => {
+                log::error!("TTS job {job_id} failed: {:?}", e);
+                finish_job(
+                    job_id,
+                    TtsJobStatus::Error,
+                    None,
+                    Some(format!("TTS failed: {e}")),
+                );
+            }
+        }
+    });
+    // Aborting an already-finished task is a no-op, so this late install is
+    // safe even if the job raced to completion.
+    with_job(job_id, |job| {
+        if !job.status.is_terminal() {
+            job.abort = Some(handle.abort_handle());
+        }
+    });
+
+    span.set_status(Status::Ok);
+    HttpResponse::Accepted().json(TtsJobCreatedResponse {
+        job_id: job_id.to_string(),
+        status: TtsJobStatus::Queued,
+    })
+}
+
+/// GET /tts/speech/jobs/{id} — poll a speech job; returns the audio once done.
+/// 404s after the job expires (results are kept ~10 min past completion).
+#[get("/tts/speech/jobs/{id}")]
+pub async fn speech_job_status_handler(
+    http_request: HttpRequest,
+    _claims: Claims,
+    path: web::Path<String>,
+) -> impl Responder {
+    let parent_context = extract_context_from_request(&http_request);
+    let mut span =
+        global_tracer().start_with_context("http.tts.speech_job.status", &parent_context);
+
+    let Ok(id) = Uuid::parse_str(&path.into_inner()) else {
+        span.set_status(Status::error("invalid job id"));
+        return HttpResponse::BadRequest().json(json!({ "error": "invalid job id" }));
+    };
+    let resp = {
+        let jobs = TTS_JOBS.lock().unwrap();
+        jobs.get(&id).map(|job| TtsJobStatusResponse {
+            job_id: id.to_string(),
+            status: job.status,
+            format: job.format.clone(),
+            audio_base64: job.audio_base64.clone(),
+            error: job.error.clone(),
+        })
+    };
+    match resp {
+        Some(r) => {
+            span.set_status(Status::Ok);
+            HttpResponse::Ok().json(r)
+        }
+        None => {
+            span.set_status(Status::error("job not found"));
+            HttpResponse::NotFound()
+                .json(json!({ "error": "TTS job not found (it may have expired)" }))
+        }
+    }
+}
+
+/// DELETE /tts/speech/jobs/{id} — cancel a queued/running speech job. Once the
+/// upstream GPU job has started it can't be interrupted (same wrapper
+/// limitation as the sync path); cancelling stops the wait and discards the
+/// result. Cancelling an already-finished job leaves it terminal.
+#[delete("/tts/speech/jobs/{id}")]
+pub async fn cancel_speech_job_handler(
+    http_request: HttpRequest,
+    _claims: Claims,
+    path: web::Path<String>,
+) -> impl Responder {
+    let parent_context = extract_context_from_request(&http_request);
+    let mut span =
+        global_tracer().start_with_context("http.tts.speech_job.cancel", &parent_context);
+
+    let Ok(id) = Uuid::parse_str(&path.into_inner()) else {
+        span.set_status(Status::error("invalid job id"));
+        return HttpResponse::BadRequest().json(json!({ "error": "invalid job id" }));
+    };
+    let status = with_job(id, |job| {
+        if !job.status.is_terminal() {
+            if let Some(h) = job.abort.take() {
+                h.abort();
+            }
+            job.status = TtsJobStatus::Cancelled;
+            job.finished_at = Some(Instant::now());
+        }
+        job.status
+    });
+    match status {
+        Some(s) => {
+            span.set_status(Status::Ok);
+            HttpResponse::Ok().json(json!({ "job_id": id.to_string(), "status": s }))
+        }
+        None => {
+            span.set_status(Status::error("job not found"));
+            HttpResponse::NotFound()
+                .json(json!({ "error": "TTS job not found (it may have expired)" }))
+        }
+    }
+}
+
+#[derive(Debug, Deserialize)]
+pub struct ListVoicesQuery {
+    /// `?refresh=1` bypasses the voice-list cache and re-queries upstream
+    /// (which may spin up the TTS model).
+    #[serde(default)]
+    pub refresh: Option<String>,
+}
+
+/// GET /tts/voices — list the Chatterbox voice library. Served from an
+/// in-memory cache when possible so browsing settings doesn't make llama-swap
+/// load the TTS model (and evict the resident LLM); see VOICES_CACHE.
 #[get("/tts/voices")]
 pub async fn list_voices_handler(
     http_request: HttpRequest,
     _claims: Claims,
+    query: web::Query<ListVoicesQuery>,
     app_state: web::Data<AppState>,
 ) -> impl Responder {
     let parent_context = extract_context_from_request(&http_request);
     let mut span = global_tracer().start_with_context("http.tts.voices.list", &parent_context);
 
+    let force = query
+        .refresh
+        .as_deref()
+        .is_some_and(|v| matches!(v, "1" | "true" | "yes"));
+    if !force && let Some(v) = cached_voices() {
+        span.set_attribute(KeyValue::new("tts.voices_cache_hit", true));
+        span.set_status(Status::Ok);
+        return HttpResponse::Ok().json(v);
+    }
+
     let Some(client) = app_state.llamacpp.as_ref() else {
         span.set_status(Status::error("tts backend not configured"));
         return HttpResponse::ServiceUnavailable()
@@ -293,6 +740,8 @@ pub async fn list_voices_handler(
     };
     match client.list_voices().await {
         Ok(v) => {
+            store_voices_cache(&v);
+            span.set_attribute(KeyValue::new("tts.voices_cache_hit", false));
             span.set_status(Status::Ok);
             HttpResponse::Ok().json(v)
         }
@@ -304,8 +753,52 @@ pub async fn list_voices_handler(
     }
 }
 
+/// DELETE /tts/voices/{name} — remove a cloned voice from the library.
+#[delete("/tts/voices/{name}")]
+pub async fn delete_voice_handler(
+    http_request: HttpRequest,
+    _claims: Claims,
+    path: web::Path<String>,
+    app_state: web::Data<AppState>,
+) -> impl Responder {
+    let parent_context = extract_context_from_request(&http_request);
+    let mut span = global_tracer().start_with_context("http.tts.voices.delete", &parent_context);
+
+    let Some(client) = app_state.llamacpp.as_ref() else {
+        span.set_status(Status::error("tts backend not configured"));
+        return HttpResponse::ServiceUnavailable()
+            .json(json!({ "error": "TTS backend not configured" }));
+    };
+    // Same charset rule as creation — a name that sanitizes differently was
+    // never a voice we created, and must not reach the upstream URL.
+    let raw = path.into_inner();
+    let name = match sanitize_voice_name(&raw) {
+        Some(n) if n == raw => n,
+        _ => {
+            span.set_status(Status::error("invalid voice name"));
+            return HttpResponse::BadRequest().json(json!({ "error": "invalid voice name" }));
+        }
+    };
+    span.set_attribute(KeyValue::new("tts.voice_name", name.clone()));
+
+    match client.delete_voice(&name).await {
+        Ok(v) => {
+            invalidate_voices_cache();
+            span.set_status(Status::Ok);
+            HttpResponse::Ok().json(v)
+        }
+        Err(e) => {
+            span.set_status(Status::error("delete_voice failed"));
+            log::error!("delete_voice failed: {:?}", e);
+            HttpResponse::BadGateway().json(json!({ "error": format!("{e}") }))
+        }
+    }
+}
+
 /// POST /tts/voices/upload — register a cloned voice from an uploaded audio
-/// clip. Multipart fields: `voice_name` (text) + a file part (`voice_file`).
+/// clip. Multipart fields: `voice_name` (text) + a file part (`voice_file`),
+/// plus optional `start_seconds` / `duration_seconds` (text) selecting which
+/// window of a longer recording becomes the reference clip.
 #[post("/tts/voices/upload")]
 pub async fn create_voice_upload_handler(
     http_request: HttpRequest,
@@ -323,6 +816,8 @@ pub async fn create_voice_upload_handler(
     };
 
     let mut voice_name: Option<String> = None;
+    let mut start_field: Option<String> = None;
+    let mut duration_field: Option<String> = None;
     let mut file_bytes = BytesMut::new();
     let mut filename = "voice.wav".to_string();
 
@@ -347,22 +842,57 @@ pub async fn create_voice_upload_handler(
                 }
                 file_bytes.put(data);
             }
-        } else if name_opt.as_deref() == Some("voice_name") {
+        } else if matches!(
+            name_opt.as_deref(),
+            Some("voice_name" | "start_seconds" | "duration_seconds")
+        ) {
+            let field = name_opt.as_deref().unwrap().to_string();
             let mut buf = BytesMut::new();
             while let Some(Ok(data)) = part.next().await {
                 buf.put(data);
             }
-            voice_name = Some(String::from_utf8_lossy(&buf).trim().to_string());
+            let text = String::from_utf8_lossy(&buf).trim().to_string();
+            match field.as_str() {
+                "voice_name" => voice_name = Some(text),
+                "start_seconds" => start_field = Some(text),
+                _ => duration_field = Some(text),
+            }
         } else {
             while let Some(Ok(_)) = part.next().await {}
         }
     }
 
+    // Empty text parts are treated as absent; anything else must parse, so a
+    // client bug ("abc") fails loudly instead of silently cloning from 0s.
+    let parse_secs = |field: Option<&String>, name: &str| -> Result<Option<f64>, String> {
+        match field.map(|s| s.as_str()).filter(|s| !s.is_empty()) {
+            None => Ok(None),
+            Some(s) => s
+                .parse::<f64>()
+                .map(Some)
+                .map_err(|_| format!("{name} must be a number of seconds")),
+        }
+    };
+    let window = parse_secs(start_field.as_ref(), "start_seconds").and_then(|start| {
+        parse_secs(duration_field.as_ref(), "duration_seconds")
+            .and_then(|dur| resolve_ref_window(start, dur))
+    });
+    let (ref_start, ref_duration) = match window {
+        Ok(w) => w,
+        Err(msg) => {
+            span.set_status(Status::error("invalid reference window"));
+            return HttpResponse::BadRequest().json(json!({ "error": msg }));
+        }
+    };
+
     let Some(name) = voice_name.as_deref().and_then(sanitize_voice_name) else {
         span.set_status(Status::error("voice_name is required"));
         return HttpResponse::BadRequest()
             .json(json!({ "error": "voice_name is required (alphanumerics, - and _ only)" }));
     };
+    // Tag the name with the ref-clip length (e.g. `grandma-30s`) so the
+    // library shows which reference length produced each clone.
+    let name = append_ref_window(&name, ref_start, ref_duration.round().max(1.0) as u32);
     if file_bytes.is_empty() {
         span.set_status(Status::error("voice_file is required"));
         return HttpResponse::BadRequest().json(json!({ "error": "voice_file is required" }));
@@ -373,21 +903,23 @@ pub async fn create_voice_upload_handler(
     // Normalize to WAV so any device format (e.g. .aac / .opus, which Chatterbox
     // rejects by extension) is accepted.
     let src_ext = Path::new(&filename).extension().and_then(|e| e.to_str());
-    let wav = match transcode_bytes_to_wav(file_bytes.as_ref(), src_ext).await {
-        Ok(w) => w,
-        Err(e) => {
-            span.set_status(Status::error("audio decode failed"));
-            log::error!("voice upload transcode failed: {:?}", e);
-            return HttpResponse::BadRequest()
-                .json(json!({ "error": "couldn't decode that audio file" }));
-        }
-    };
+    let wav =
+        match transcode_bytes_to_wav(file_bytes.as_ref(), src_ext, ref_start, ref_duration).await {
+            Ok(w) => w,
+            Err(e) => {
+                span.set_status(Status::error("audio decode failed"));
+                log::error!("voice upload transcode failed: {:?}", e);
+                return HttpResponse::BadRequest()
+                    .json(json!({ "error": "couldn't decode that audio file" }));
+            }
+        };
 
     match client
         .create_voice(&name, wav, "reference.wav", "audio/wav")
         .await
     {
         Ok(v) => {
+            invalidate_voices_cache();
             span.set_status(Status::Ok);
             HttpResponse::Ok().json(v)
         }
@@ -406,11 +938,19 @@ pub struct CreateVoiceFromLibraryRequest {
     pub path: String,
     #[serde(default)]
     pub library: Option<String>,
+    /// Offset into the source where the reference window begins (default 0) —
+    /// lets the client pick the clean-speech section of a long recording.
+    #[serde(default)]
+    pub start_seconds: Option<f64>,
+    /// Reference window length; clamped to LLAMA_SWAP_TTS_REF_SECONDS.
+    #[serde(default)]
+    pub duration_seconds: Option<f64>,
 }
 
 /// POST /tts/voices/from-library — register a cloned voice from a file already
 /// in a library. Audio and video alike are ffmpeg-normalized to a mono 24 kHz
-/// WAV reference clip (length capped by LLAMA_SWAP_TTS_REF_SECONDS).
+/// WAV reference clip (window selected by start/duration_seconds, length
+/// capped by LLAMA_SWAP_TTS_REF_SECONDS).
 #[post("/tts/voices/from-library")]
 pub async fn create_voice_from_library_handler(
     http_request: HttpRequest,
@@ -432,6 +972,18 @@ pub async fn create_voice_from_library_handler(
         return HttpResponse::BadRequest()
             .json(json!({ "error": "voice_name is required (alphanumerics, - and _ only)" }));
     };
+    let (ref_start, ref_duration) =
+        match resolve_ref_window(req.start_seconds, req.duration_seconds) {
+            Ok(w) => w,
+            Err(msg) => {
+                span.set_status(Status::error("invalid reference window"));
+                return HttpResponse::BadRequest().json(json!({ "error": msg }));
+            }
+        };
+    // Tag the name with the ref-clip length (e.g. `grandma-30s`) so the
+    // library shows which reference length produced each clone.
+    let voice_name =
+        append_ref_window(&voice_name, ref_start, ref_duration.round().max(1.0) as u32);
 
     let library = match libraries::resolve_library_param(&app_state, req.library.as_deref()) {
         Ok(Some(l)) => l,
@@ -460,7 +1012,7 @@ pub async fn create_voice_from_library_handler(
     }
     span.set_attribute(KeyValue::new("tts.voice_name", voice_name.clone()));
 
-    let wav = match prepare_reference_audio(&abs).await {
+    let wav = match prepare_reference_audio(&abs, ref_start, ref_duration).await {
         Ok(b) => b,
         Err(e) => {
             span.set_status(Status::error("audio decode failed"));
@@ -475,6 +1027,7 @@ pub async fn create_voice_from_library_handler(
         .await
     {
         Ok(v) => {
+            invalidate_voices_cache();
             span.set_status(Status::Ok);
             HttpResponse::Ok().json(v)
         }
@@ -489,8 +1042,8 @@ pub async fn create_voice_from_library_handler(
 /// Read a library file (audio or video) as a Chatterbox-ready reference: ffmpeg
 /// decodes/extracts its audio to mono 24 kHz WAV. Reading straight from the
 /// library path avoids slurping a (possibly large) video into memory.
-async fn prepare_reference_audio(abs: &Path) -> anyhow::Result<Vec<u8>> {
-    run_ffmpeg_to_wav(&abs.to_string_lossy()).await
+async fn prepare_reference_audio(abs: &Path, start: f64, duration: f64) -> anyhow::Result<Vec<u8>> {
+    run_ffmpeg_to_wav(&abs.to_string_lossy(), start, duration).await
 }
 
 #[cfg(test)]
@@ -534,6 +1087,151 @@ mod tests {
         assert_eq!(sanitize_voice_name(&long).unwrap().len(), 64);
     }
 
+    #[test]
+    fn append_ref_window_tags_name() {
+        assert_eq!(append_ref_window("grandma", 0.0, 30), "grandma-30s");
+        assert_eq!(append_ref_window("voice_01", 0.0, 15), "voice_01-15s");
+    }
+
+    #[test]
+    fn append_ref_window_includes_nonzero_start() {
+        // Sub-minute starts stay in seconds; longer ones read as XmYYs since
+        // ':' isn't allowed in voice names.
+        assert_eq!(append_ref_window("grandma", 45.0, 30), "grandma-at45s-30s");
+        assert_eq!(
+            append_ref_window("grandma", 92.4, 30),
+            "grandma-at1m32s-30s"
+        );
+        assert_eq!(
+            append_ref_window("grandma", 600.0, 12),
+            "grandma-at10m00s-12s"
+        );
+        // A start that rounds to zero is "from the start".
+        assert_eq!(append_ref_window("grandma", 0.3, 30), "grandma-30s");
+    }
+
+    #[test]
+    fn append_ref_window_is_idempotent_for_same_window() {
+        assert_eq!(append_ref_window("grandma-30s", 0.0, 30), "grandma-30s");
+        assert_eq!(
+            append_ref_window("grandma-at45s-30s", 45.0, 30),
+            "grandma-at45s-30s"
+        );
+        // A different window still appends — that's the comparison use-case.
+        assert_eq!(append_ref_window("grandma-15s", 0.0, 30), "grandma-15s-30s");
+        assert_eq!(
+            append_ref_window("grandma-30s", 45.0, 30),
+            "grandma-30s-at45s-30s"
+        );
+    }
+
+    #[test]
+    fn append_ref_window_keeps_64_char_bound() {
+        let long = "a".repeat(64);
+        let tagged = append_ref_window(&long, 0.0, 30);
+        assert_eq!(tagged.len(), 64);
+        assert!(tagged.ends_with("-30s"));
+
+        let tagged = append_ref_window(&long, 92.0, 30);
+        assert_eq!(tagged.len(), 64);
+        assert!(tagged.ends_with("-at1m32s-30s"));
+    }
+
+    #[test]
+    fn resolve_ref_window_defaults_to_start_of_clip_at_cap_length() {
+        // Reads the live cap rather than mutating LLAMA_SWAP_TTS_REF_SECONDS:
+        // env mutation flakes under the parallel suite (see env_dispatch).
+        let cap = f64::from(tts_ref_seconds());
+        assert_eq!(resolve_ref_window(None, None), Ok((0.0, cap)));
+    }
+
+    #[test]
+    fn resolve_ref_window_accepts_offset_and_clamps_duration() {
+        let cap = f64::from(tts_ref_seconds());
+        assert_eq!(resolve_ref_window(Some(92.5), None), Ok((92.5, cap)));
+        assert_eq!(resolve_ref_window(Some(10.0), Some(12.0)), Ok((10.0, 12.0)));
+        // Longer-than-cap windows are bounded, not rejected.
+        assert_eq!(resolve_ref_window(None, Some(cap + 100.0)), Ok((0.0, cap)));
+    }
+
+    #[test]
+    fn resolve_ref_window_rejects_garbage() {
+        assert!(resolve_ref_window(Some(-1.0), None).is_err());
+        assert!(resolve_ref_window(Some(f64::NAN), None).is_err());
+        assert!(resolve_ref_window(Some(f64::INFINITY), None).is_err());
+        assert!(resolve_ref_window(None, Some(0.0)).is_err());
+        assert!(resolve_ref_window(None, Some(-5.0)).is_err());
+        assert!(resolve_ref_window(None, Some(f64::NAN)).is_err());
+    }
+
+    #[test]
+    fn sweep_drops_expired_results_and_keeps_live_jobs() {
+        let now = Instant::now();
+        let mk = |status: TtsJobStatus, created: Instant, finished: Option<Instant>| TtsJob {
+            status,
+            format: "mp3".into(),
+            audio_base64: None,
+            error: None,
+            created_at: created,
+            finished_at: finished,
+            abort: None,
+        };
+        let mut jobs = HashMap::new();
+        let live = Uuid::new_v4();
+        let fresh_done = Uuid::new_v4();
+        let stale_done = Uuid::new_v4();
+        jobs.insert(live, mk(TtsJobStatus::Running, now, None));
+        jobs.insert(
+            fresh_done,
+            mk(TtsJobStatus::Done, now, Some(now - Duration::from_secs(60))),
+        );
+        jobs.insert(
+            stale_done,
+            mk(
+                TtsJobStatus::Done,
+                now - TTS_JOB_MAX_AGE / 2,
+                Some(now - TTS_JOB_RESULT_TTL),
+            ),
+        );
+
+        sweep_stale_jobs(&mut jobs, now);
+        assert!(jobs.contains_key(&live));
+        assert!(jobs.contains_key(&fresh_done));
+        assert!(!jobs.contains_key(&stale_done));
+    }
+
+    #[test]
+    fn sweep_drops_jobs_past_max_age_even_if_unfinished() {
+        let now = Instant::now();
+        let mut jobs = HashMap::new();
+        let ancient = Uuid::new_v4();
+        jobs.insert(
+            ancient,
+            TtsJob {
+                status: TtsJobStatus::Running,
+                format: "mp3".into(),
+                audio_base64: None,
+                error: None,
+                created_at: now - TTS_JOB_MAX_AGE,
+                finished_at: None,
+                abort: None,
+            },
+        );
+        sweep_stale_jobs(&mut jobs, now);
+        assert!(jobs.is_empty());
+    }
+
+    #[test]
+    fn voices_cache_roundtrip_and_invalidation() {
+        invalidate_voices_cache();
+        assert!(cached_voices().is_none());
+        let v = json!({ "voices": [{ "name": "m-30s" }], "count": 1 });
+        store_voices_cache(&v);
+        assert_eq!(cached_voices(), Some(v));
+        invalidate_voices_cache();
+        assert!(cached_voices().is_none());
+    }
+
     #[test]
     fn clean_for_tts_strips_markdown() {
         assert_eq!(

src/bin/import_calendar.rs

@@ -1,7 +1,7 @@
 use anyhow::{Context, Result};
 use chrono::Utc;
 use clap::Parser;
-use image_api::ai::ollama::OllamaClient;
+use image_api::ai::LocalLlm;
 use image_api::bin_progress;
 use image_api::database::calendar_dao::{InsertCalendarEvent, SqliteCalendarEventDao};
 use image_api::parsers::ical_parser::parse_ics_file;
@@ -44,22 +44,10 @@ async fn main() -> Result<()> {
 
     let context = opentelemetry::Context::current();
 
-    let ollama = if args.generate_embeddings {
-        let primary_url = dotenv::var("OLLAMA_PRIMARY_URL")
-            .or_else(|_| dotenv::var("OLLAMA_URL"))
-            .unwrap_or_else(|_| "http://localhost:11434".to_string());
-        let fallback_url = dotenv::var("OLLAMA_FALLBACK_URL").ok();
-        let primary_model = dotenv::var("OLLAMA_PRIMARY_MODEL")
-            .or_else(|_| dotenv::var("OLLAMA_MODEL"))
-            .unwrap_or_else(|_| "nomic-embed-text:v1.5".to_string());
-        let fallback_model = dotenv::var("OLLAMA_FALLBACK_MODEL").ok();
-
-        Some(OllamaClient::new(
-            primary_url,
-            fallback_url,
-            primary_model,
-            fallback_model,
-        ))
+    // LocalLlm dispatches per LLM_BACKEND, so embeddings written here land
+    // in the same vector space the query side searches.
+    let llm = if args.generate_embeddings {
+        Some(LocalLlm::from_env())
     } else {
         None
     };
@@ -90,7 +78,7 @@ async fn main() -> Result<()> {
         }
 
         // Generate embedding if requested (blocking call)
-        let embedding = if let Some(ref ollama_client) = ollama {
+        let embedding = if let Some(ref llm) = llm {
             let text = format!(
                 "{} {} {}",
                 event.summary,
@@ -100,7 +88,7 @@ async fn main() -> Result<()> {
 
             match tokio::task::block_in_place(|| {
                 tokio::runtime::Handle::current()
-                    .block_on(async { ollama_client.generate_embedding(&text).await })
+                    .block_on(async { llm.embed_document(&text).await })
             }) {
                 Ok(emb) => Some(emb),
                 Err(e) => {

src/bin/import_search_history.rs

@@ -1,7 +1,7 @@
 use anyhow::{Context, Result};
 use chrono::Utc;
 use clap::Parser;
-use image_api::ai::ollama::OllamaClient;
+use image_api::ai::LocalLlm;
 use image_api::bin_progress;
 use image_api::database::search_dao::{InsertSearchRecord, SqliteSearchHistoryDao};
 use image_api::parsers::search_html_parser::parse_search_html;
@@ -38,16 +38,9 @@ async fn main() -> Result<()> {
 
     info!("Found {} search records", searches.len());
 
-    let primary_url = dotenv::var("OLLAMA_PRIMARY_URL")
-        .or_else(|_| dotenv::var("OLLAMA_URL"))
-        .unwrap_or_else(|_| "http://localhost:11434".to_string());
-    let fallback_url = dotenv::var("OLLAMA_FALLBACK_URL").ok();
-    let primary_model = dotenv::var("OLLAMA_PRIMARY_MODEL")
-        .or_else(|_| dotenv::var("OLLAMA_MODEL"))
-        .unwrap_or_else(|_| "nomic-embed-text:v1.5".to_string());
-    let fallback_model = dotenv::var("OLLAMA_FALLBACK_MODEL").ok();
-
-    let ollama = OllamaClient::new(primary_url, fallback_url, primary_model, fallback_model);
+    // LocalLlm dispatches per LLM_BACKEND, so embeddings written here land
+    // in the same vector space the query side searches.
+    let llm = LocalLlm::from_env();
     let context = opentelemetry::Context::current();
 
     let mut inserted_count = 0usize;
@@ -67,12 +60,11 @@ async fn main() -> Result<()> {
 
         let pb_for_warn = pb.clone();
         let embeddings_result = tokio::task::spawn({
-            let ollama_client = ollama.clone();
+            let llm = llm.clone();
             async move {
-                // Generate embeddings in parallel for the batch
                 let mut embeddings = Vec::new();
                 for query in &queries {
-                    match ollama_client.generate_embedding(query).await {
+                    match llm.embed_document(query).await {
                         Ok(emb) => embeddings.push(Some(emb)),
                         Err(e) => {
                             pb_for_warn.println(format!("embedding failed for '{}': {}", query, e));

src/bin/reembed_embeddings.rs

@@ -0,0 +1,465 @@
+//! Re-embed stored corpora through `LocalLlm`, i.e. the same
+//! `LLM_BACKEND` dispatch the query side uses. The original import /
+//! backfill tools always embedded via Ollama, so a deploy running
+//! `LLM_BACKEND=llamacpp` queries vector spaces the corpora may not live
+//! in. Three tables share the problem and are all covered here:
+//!
+//! - `daily_conversation_summaries` — re-embeds
+//!   `strip_summary_boilerplate(summary)` (what the original job fed the
+//!   embedder); also rewrites `model_version`.
+//! - `calendar_events` — re-embeds "summary description location" exactly
+//!   as `import_calendar` does; rows without an embedding are skipped (the
+//!   import only embeds under `--generate-embeddings`).
+//! - `search_history` — re-embeds the raw query text.
+//! - `entities` (knowledge graph) — re-embeds "name description" exactly as
+//!   `tool_store_entity` does; embedding-less rows are skipped (embedding
+//!   is best-effort at store time).
+//!
+//! Source text is untouched — only vectors are rewritten. The old↔new
+//! cosine report doubles as a diagnostic: ~1.0 means both backends already
+//! shared a space (re-embedding was a no-op); low values confirm the
+//! mismatch this tool exists to fix.
+
+use anyhow::{Context, Result};
+use clap::Parser;
+use diesel::prelude::*;
+use diesel::sql_query;
+use diesel::sqlite::SqliteConnection;
+use image_api::ai::{LocalLlm, strip_summary_boilerplate};
+use image_api::bin_progress;
+use std::env;
+
+#[derive(Parser, Debug)]
+#[command(author, version, about = "Re-embed stored corpora via the configured LLM_BACKEND", long_about = None)]
+struct Args {
+    /// Comma-separated tables to process: summaries, calendar, search, entities
+    #[arg(long, default_value = "summaries,calendar,search,entities")]
+    tables: String,
+
+    /// Only process the first N rows per table (smoke test)
+    #[arg(long)]
+    limit: Option<usize>,
+
+    /// Compute embeddings and report old↔new similarity without writing
+    #[arg(long, default_value_t = false)]
+    dry_run: bool,
+}
+
+#[derive(QueryableByName)]
+struct SummaryRow {
+    #[diesel(sql_type = diesel::sql_types::Integer)]
+    id: i32,
+    #[diesel(sql_type = diesel::sql_types::Text)]
+    summary: String,
+    #[diesel(sql_type = diesel::sql_types::Binary)]
+    embedding: Vec<u8>,
+    #[diesel(sql_type = diesel::sql_types::Text)]
+    model_version: String,
+}
+
+#[derive(QueryableByName)]
+struct CalendarRow {
+    #[diesel(sql_type = diesel::sql_types::Integer)]
+    id: i32,
+    #[diesel(sql_type = diesel::sql_types::Text)]
+    summary: String,
+    #[diesel(sql_type = diesel::sql_types::Nullable<diesel::sql_types::Text>)]
+    description: Option<String>,
+    #[diesel(sql_type = diesel::sql_types::Nullable<diesel::sql_types::Text>)]
+    location: Option<String>,
+    #[diesel(sql_type = diesel::sql_types::Binary)]
+    embedding: Vec<u8>,
+}
+
+#[derive(QueryableByName)]
+struct SearchRow {
+    #[diesel(sql_type = diesel::sql_types::BigInt)]
+    id: i64,
+    #[diesel(sql_type = diesel::sql_types::Text)]
+    query: String,
+    #[diesel(sql_type = diesel::sql_types::Binary)]
+    embedding: Vec<u8>,
+}
+
+#[derive(QueryableByName)]
+struct EntityRow {
+    #[diesel(sql_type = diesel::sql_types::Integer)]
+    id: i32,
+    #[diesel(sql_type = diesel::sql_types::Text)]
+    name: String,
+    #[diesel(sql_type = diesel::sql_types::Text)]
+    description: String,
+    #[diesel(sql_type = diesel::sql_types::Binary)]
+    embedding: Vec<u8>,
+}
+
+/// One unit of re-embed work, normalized across tables.
+struct WorkItem {
+    /// Row key, as i64 so both i32 ids and rowids fit.
+    id: i64,
+    /// Text fed to the embedder — must match what the original writer used.
+    text: String,
+    /// Existing vector bytes, for the old↔new similarity report.
+    old_embedding: Vec<u8>,
+}
+
+fn deserialize_vector(bytes: &[u8]) -> Option<Vec<f32>> {
+    if !bytes.len().is_multiple_of(4) {
+        return None;
+    }
+    Some(
+        bytes
+            .chunks_exact(4)
+            .map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]]))
+            .collect(),
+    )
+}
+
+fn serialize_vector(vec: &[f32]) -> Vec<u8> {
+    vec.iter().flat_map(|f| f.to_le_bytes()).collect()
+}
+
+fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
+    if a.len() != b.len() {
+        return 0.0;
+    }
+    let dot: f32 = a.iter().zip(b).map(|(x, y)| x * y).sum();
+    let mag_a: f32 = a.iter().map(|x| x * x).sum::<f32>().sqrt();
+    let mag_b: f32 = b.iter().map(|x| x * x).sum::<f32>().sqrt();
+    if mag_a == 0.0 || mag_b == 0.0 {
+        return 0.0;
+    }
+    dot / (mag_a * mag_b)
+}
+
+/// Embed `text`, halving it on "input too large" errors until it fits the
+/// server's physical batch (`--ubatch-size`). Mirrors the silent truncation
+/// Ollama applied when these corpora were first embedded — llama-server
+/// returns a 500 instead — except here it's surfaced via the returned flag.
+/// Returns `(embedding, truncated)`.
+async fn embed_with_truncation(llm: &LocalLlm, text: &str) -> Result<(Vec<f32>, bool)> {
+    let mut text = text.to_string();
+    let mut truncated = false;
+    loop {
+        match llm.embed_document(&text).await {
+            Ok(emb) => return Ok((emb, truncated)),
+            Err(e)
+                if e.to_string().contains("too large to process") && text.chars().count() > 64 =>
+            {
+                let keep = text.chars().count() / 2;
+                text = text.chars().take(keep).collect();
+                truncated = true;
+            }
+            Err(e) => return Err(e),
+        }
+    }
+}
+
+/// Re-embed `items`, writing each new vector via `update`. Returns the
+/// old↔new cosines for the similarity report.
+async fn reembed_table(
+    conn: &mut SqliteConnection,
+    llm: &LocalLlm,
+    label: &str,
+    items: Vec<WorkItem>,
+    dry_run: bool,
+    update: impl Fn(&mut SqliteConnection, i64, Vec<u8>) -> Result<()>,
+) -> Result<Vec<f32>> {
+    println!("\n[{}] re-embedding {} rows...", label, items.len());
+    let pb = bin_progress::determinate(items.len() as u64, format!("re-embedding {}", label));
+
+    let mut sims: Vec<f32> = Vec::with_capacity(items.len());
+    let mut updated = 0usize;
+    let mut failed = 0usize;
+    let mut truncated_count = 0usize;
+
+    for item in &items {
+        let new_emb = match embed_with_truncation(llm, &item.text).await {
+            Ok((e, truncated)) => {
+                if truncated {
+                    truncated_count += 1;
+                    pb.println(format!(
+                        "⚠ {} id={}: input exceeded the embed server's batch size, \
+                         truncated before embedding",
+                        label, item.id
+                    ));
+                }
+                e
+            }
+            Err(e) => {
+                pb.inc(1);
+                failed += 1;
+                eprintln!("✗ {} id={}: {}", label, item.id, e);
+                continue;
+            }
+        };
+
+        // The whole pipeline (DAO checks, stored corpora) assumes
+        // EMBEDDING_DIM dims. A mismatch means the active embed slot is not
+        // serving the configured model — stop rather than corrupt the table.
+        anyhow::ensure!(
+            new_emb.len() == image_api::ai::embedding_dim(),
+            "backend returned {}-dim embedding (expected {}) — '{}' does not \
+             match the configured EMBEDDING_DIM",
+            new_emb.len(),
+            image_api::ai::embedding_dim(),
+            llm.embedding_model_version()
+        );
+
+        if let Some(old_emb) = deserialize_vector(&item.old_embedding) {
+            sims.push(cosine_similarity(&old_emb, &new_emb));
+        }
+
+        if !dry_run {
+            update(conn, item.id, serialize_vector(&new_emb))
+                .with_context(|| format!("updating {} id={}", label, item.id))?;
+        }
+        updated += 1;
+        pb.inc(1);
+    }
+    pb.finish_and_clear();
+
+    println!(
+        "[{}] {} re-embedded ({} truncated), {} failed",
+        label, updated, truncated_count, failed
+    );
+    Ok(sims)
+}
+
+fn report_similarity(label: &str, mut sims: Vec<f32>) {
+    if sims.is_empty() {
+        println!("[{}] no old↔new pairs to compare", label);
+        return;
+    }
+    sims.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    let mean: f32 = sims.iter().sum::<f32>() / sims.len() as f32;
+    let median = sims[sims.len() / 2];
+    println!(
+        "[{}] old↔new cosine over identical text: min={:.3} median={:.3} mean={:.3} max={:.3}",
+        label,
+        sims.first().unwrap(),
+        median,
+        mean,
+        sims.last().unwrap()
+    );
+    if median > 0.98 {
+        println!(
+            "[{}] → old and new backends agree (~same vector space); poor search \
+             results are coming from something else (prefixes, thresholds, corpus).",
+            label
+        );
+    } else if median > 0.9 {
+        println!(
+            "[{}] → same model family but measurably different vectors \
+             (quantization / runtime drift); re-embedding was worthwhile.",
+            label
+        );
+    } else {
+        println!(
+            "[{}] → vector-space mismatch confirmed — queries were searching a \
+             different space than the corpus. This re-embed should fix it.",
+            label
+        );
+    }
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    dotenv::dotenv().ok();
+    env_logger::init();
+    let args = Args::parse();
+
+    let tables: Vec<&str> = args.tables.split(',').map(|t| t.trim()).collect();
+    for t in &tables {
+        anyhow::ensure!(
+            matches!(*t, "summaries" | "calendar" | "search" | "entities"),
+            "unknown table '{}' — expected summaries, calendar, search, entities",
+            t
+        );
+    }
+
+    let database_url = env::var("DATABASE_URL").unwrap_or_else(|_| "auth.db".to_string());
+    println!("Database: {}", database_url);
+
+    let mut conn = SqliteConnection::establish(&database_url)
+        .with_context(|| format!("connecting to {}", database_url))?;
+
+    let llm = LocalLlm::from_env();
+    let model_version = llm.embedding_model_version();
+    println!("Embedding via '{}'", model_version);
+    if args.dry_run {
+        println!("DRY RUN — no rows will be written");
+    }
+
+    if tables.contains(&"summaries") {
+        let mut rows: Vec<SummaryRow> = sql_query(
+            "SELECT id, summary, embedding, model_version
+             FROM daily_conversation_summaries ORDER BY date",
+        )
+        .load(&mut conn)
+        .context("loading daily summaries")?;
+        if let Some(limit) = args.limit {
+            rows.truncate(limit);
+        }
+        if let Some(first) = rows.first() {
+            println!(
+                "\n[summaries] previous model_version '{}' → '{}'",
+                first.model_version, model_version
+            );
+        }
+        let items = rows
+            .into_iter()
+            .map(|r| WorkItem {
+                id: r.id as i64,
+                text: strip_summary_boilerplate(&r.summary),
+                old_embedding: r.embedding,
+            })
+            .collect();
+        let mv = model_version.clone();
+        let sims = reembed_table(
+            &mut conn,
+            &llm,
+            "summaries",
+            items,
+            args.dry_run,
+            move |conn, id, emb| {
+                sql_query(
+                    "UPDATE daily_conversation_summaries
+                     SET embedding = ?1, model_version = ?2 WHERE id = ?3",
+                )
+                .bind::<diesel::sql_types::Binary, _>(emb)
+                .bind::<diesel::sql_types::Text, _>(&mv)
+                .bind::<diesel::sql_types::Integer, _>(id as i32)
+                .execute(conn)?;
+                Ok(())
+            },
+        )
+        .await?;
+        report_similarity("summaries", sims);
+    }
+
+    if tables.contains(&"calendar") {
+        let mut rows: Vec<CalendarRow> = sql_query(
+            "SELECT id, summary, description, location, embedding
+             FROM calendar_events WHERE embedding IS NOT NULL ORDER BY id",
+        )
+        .load(&mut conn)
+        .context("loading calendar events")?;
+        if let Some(limit) = args.limit {
+            rows.truncate(limit);
+        }
+        let items = rows
+            .into_iter()
+            .map(|r| WorkItem {
+                id: r.id as i64,
+                // Same text construction as import_calendar.
+                text: format!(
+                    "{} {} {}",
+                    r.summary,
+                    r.description.as_deref().unwrap_or(""),
+                    r.location.as_deref().unwrap_or("")
+                ),
+                old_embedding: r.embedding,
+            })
+            .collect();
+        let sims = reembed_table(
+            &mut conn,
+            &llm,
+            "calendar",
+            items,
+            args.dry_run,
+            |conn, id, emb| {
+                sql_query("UPDATE calendar_events SET embedding = ?1 WHERE id = ?2")
+                    .bind::<diesel::sql_types::Binary, _>(emb)
+                    .bind::<diesel::sql_types::Integer, _>(id as i32)
+                    .execute(conn)?;
+                Ok(())
+            },
+        )
+        .await?;
+        report_similarity("calendar", sims);
+    }
+
+    if tables.contains(&"search") {
+        let mut rows: Vec<SearchRow> = sql_query(
+            "SELECT rowid AS id, query, embedding
+             FROM search_history ORDER BY rowid",
+        )
+        .load(&mut conn)
+        .context("loading search history")?;
+        if let Some(limit) = args.limit {
+            rows.truncate(limit);
+        }
+        let items = rows
+            .into_iter()
+            .map(|r| WorkItem {
+                id: r.id,
+                text: r.query,
+                old_embedding: r.embedding,
+            })
+            .collect();
+        let sims = reembed_table(
+            &mut conn,
+            &llm,
+            "search",
+            items,
+            args.dry_run,
+            |conn, id, emb| {
+                sql_query("UPDATE search_history SET embedding = ?1 WHERE rowid = ?2")
+                    .bind::<diesel::sql_types::Binary, _>(emb)
+                    .bind::<diesel::sql_types::BigInt, _>(id)
+                    .execute(conn)?;
+                Ok(())
+            },
+        )
+        .await?;
+        report_similarity("search", sims);
+    }
+
+    if tables.contains(&"entities") {
+        let mut rows: Vec<EntityRow> = sql_query(
+            "SELECT id, name, description, embedding
+             FROM entities WHERE embedding IS NOT NULL ORDER BY id",
+        )
+        .load(&mut conn)
+        .context("loading knowledge entities")?;
+        if let Some(limit) = args.limit {
+            rows.truncate(limit);
+        }
+        let items = rows
+            .into_iter()
+            .map(|r| WorkItem {
+                id: r.id as i64,
+                // Same text construction as tool_store_entity.
+                text: format!("{} {}", r.name, r.description),
+                old_embedding: r.embedding,
+            })
+            .collect();
+        let sims = reembed_table(
+            &mut conn,
+            &llm,
+            "entities",
+            items,
+            args.dry_run,
+            |conn, id, emb| {
+                sql_query("UPDATE entities SET embedding = ?1 WHERE id = ?2")
+                    .bind::<diesel::sql_types::Binary, _>(emb)
+                    .bind::<diesel::sql_types::Integer, _>(id as i32)
+                    .execute(conn)?;
+                Ok(())
+            },
+        )
+        .await?;
+        report_similarity("entities", sims);
+    }
+
+    println!(
+        "\n{}",
+        if args.dry_run {
+            "Dry run complete"
+        } else {
+            "Done"
+        }
+    );
+    Ok(())
+}

src/database/calendar_dao.rs

@@ -222,11 +222,12 @@ impl CalendarEventDao for SqliteCalendarEventDao {
 
             // Validate embedding dimensions if provided
             if let Some(ref emb) = event.embedding
-                && emb.len() != 768
+                && emb.len() != crate::ai::embedding_dim()
             {
                 return Err(anyhow::anyhow!(
-                    "Invalid embedding dimensions: {} (expected 768)",
-                    emb.len()
+                    "Invalid embedding dimensions: {} (expected {})",
+                    emb.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 
@@ -293,7 +294,7 @@ impl CalendarEventDao for SqliteCalendarEventDao {
                 for event in events {
                     // Validate embedding if provided
                     if let Some(ref emb) = event.embedding
-                        && emb.len() != 768
+                        && emb.len() != crate::ai::embedding_dim()
                     {
                         log::warn!(
                             "Skipping event with invalid embedding dimensions: {}",
@@ -385,10 +386,11 @@ impl CalendarEventDao for SqliteCalendarEventDao {
         trace_db_call(context, "query", "find_similar_events", |_span| {
             let mut conn = self.connection.lock().expect("Unable to get CalendarEventDao");
 
-            if query_embedding.len() != 768 {
+            if query_embedding.len() != crate::ai::embedding_dim() {
                 return Err(anyhow::anyhow!(
-                    "Invalid query embedding dimensions: {} (expected 768)",
-                    query_embedding.len()
+                    "Invalid query embedding dimensions: {} (expected {})",
+                    query_embedding.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 
@@ -461,10 +463,11 @@ impl CalendarEventDao for SqliteCalendarEventDao {
 
             // Step 2: If query embedding provided, rank by semantic similarity
             if let Some(query_emb) = query_embedding {
-                if query_emb.len() != 768 {
+                if query_emb.len() != crate::ai::embedding_dim() {
                     return Err(anyhow::anyhow!(
-                        "Invalid query embedding dimensions: {} (expected 768)",
-                        query_emb.len()
+                        "Invalid query embedding dimensions: {} (expected {})",
+                        query_emb.len(),
+                        crate::ai::embedding_dim()
                     ));
                 }
 

src/database/daily_summary_dao.rs

@@ -150,10 +150,11 @@ impl DailySummaryDao for SqliteDailySummaryDao {
                 .expect("Unable to get DailySummaryDao");
 
             // Validate embedding dimensions
-            if summary.embedding.len() != 768 {
+            if summary.embedding.len() != crate::ai::embedding_dim() {
                 return Err(anyhow::anyhow!(
-                    "Invalid embedding dimensions: {} (expected 768)",
-                    summary.embedding.len()
+                    "Invalid embedding dimensions: {} (expected {})",
+                    summary.embedding.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 
@@ -202,10 +203,11 @@ impl DailySummaryDao for SqliteDailySummaryDao {
         trace_db_call(context, "query", "find_similar_summaries", |_span| {
             let mut conn = self.connection.lock().expect("Unable to get DailySummaryDao");
 
-            if query_embedding.len() != 768 {
+            if query_embedding.len() != crate::ai::embedding_dim() {
                 return Err(anyhow::anyhow!(
-                    "Invalid query embedding dimensions: {} (expected 768)",
-                    query_embedding.len()
+                    "Invalid query embedding dimensions: {} (expected {})",
+                    query_embedding.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 
@@ -299,10 +301,11 @@ impl DailySummaryDao for SqliteDailySummaryDao {
         trace_db_call(context, "query", "find_similar_summaries_with_time_weight", |_span| {
             let mut conn = self.connection.lock().expect("Unable to get DailySummaryDao");
 
-            if query_embedding.len() != 768 {
+            if query_embedding.len() != crate::ai::embedding_dim() {
                 return Err(anyhow::anyhow!(
-                    "Invalid query embedding dimensions: {} (expected 768)",
-                    query_embedding.len()
+                    "Invalid query embedding dimensions: {} (expected {})",
+                    query_embedding.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 

src/database/location_dao.rs

@@ -216,11 +216,12 @@ impl LocationHistoryDao for SqliteLocationHistoryDao {
 
             // Validate embedding dimensions if provided (rare for location data)
             if let Some(ref emb) = location.embedding
-                && emb.len() != 768
+                && emb.len() != crate::ai::embedding_dim()
             {
                 return Err(anyhow::anyhow!(
-                    "Invalid embedding dimensions: {} (expected 768)",
-                    emb.len()
+                    "Invalid embedding dimensions: {} (expected {})",
+                    emb.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 
@@ -292,7 +293,7 @@ impl LocationHistoryDao for SqliteLocationHistoryDao {
                 for location in locations {
                     // Validate embedding if provided (rare)
                     if let Some(ref emb) = location.embedding
-                        && emb.len() != 768
+                        && emb.len() != crate::ai::embedding_dim()
                     {
                         log::warn!(
                             "Skipping location with invalid embedding dimensions: {}",

src/database/search_dao.rs

@@ -189,10 +189,11 @@ impl SearchHistoryDao for SqliteSearchHistoryDao {
                 .expect("Unable to get SearchHistoryDao");
 
             // Validate embedding dimensions (REQUIRED for searches)
-            if search.embedding.len() != 768 {
+            if search.embedding.len() != crate::ai::embedding_dim() {
                 return Err(anyhow::anyhow!(
-                    "Invalid embedding dimensions: {} (expected 768)",
-                    search.embedding.len()
+                    "Invalid embedding dimensions: {} (expected {})",
+                    search.embedding.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 
@@ -245,7 +246,7 @@ impl SearchHistoryDao for SqliteSearchHistoryDao {
             conn.transaction::<_, anyhow::Error, _>(|conn| {
                 for search in searches {
                     // Validate embedding (REQUIRED)
-                    if search.embedding.len() != 768 {
+                    if search.embedding.len() != crate::ai::embedding_dim() {
                         log::warn!(
                             "Skipping search with invalid embedding dimensions: {}",
                             search.embedding.len()
@@ -325,10 +326,11 @@ impl SearchHistoryDao for SqliteSearchHistoryDao {
                 .lock()
                 .expect("Unable to get SearchHistoryDao");
 
-            if query_embedding.len() != 768 {
+            if query_embedding.len() != crate::ai::embedding_dim() {
                 return Err(anyhow::anyhow!(
-                    "Invalid query embedding dimensions: {} (expected 768)",
-                    query_embedding.len()
+                    "Invalid query embedding dimensions: {} (expected {})",
+                    query_embedding.len(),
+                    crate::ai::embedding_dim()
                 ));
             }
 
@@ -406,10 +408,11 @@ impl SearchHistoryDao for SqliteSearchHistoryDao {
 
             // Step 2: If query embedding provided, rank by semantic similarity
             if let Some(query_emb) = query_embedding {
-                if query_emb.len() != 768 {
+                if query_emb.len() != crate::ai::embedding_dim() {
                     return Err(anyhow::anyhow!(
-                        "Invalid query embedding dimensions: {} (expected 768)",
-                        query_emb.len()
+                        "Invalid query embedding dimensions: {} (expected {})",
+                        query_emb.len(),
+                        crate::ai::embedding_dim()
                     ));
                 }
 

src/main.rs

@@ -364,9 +364,13 @@ fn main() -> std::io::Result<()> {
                 .service(ai::rate_insight_handler)
                 .service(ai::export_training_data_handler)
                 .service(ai::tts_speech_handler)
+                .service(ai::create_speech_job_handler)
+                .service(ai::speech_job_status_handler)
+                .service(ai::cancel_speech_job_handler)
                 .service(ai::list_voices_handler)
                 .service(ai::create_voice_upload_handler)
                 .service(ai::create_voice_from_library_handler)
+                .service(ai::delete_voice_handler)
                 .service(libraries::list_libraries)
                 .service(libraries::patch_library)
                 .add_feature(add_tag_services::<_, SqliteTagDao>)

src/state.rs

@@ -186,21 +186,7 @@ impl AppState {
 impl Default for AppState {
     fn default() -> Self {
         // Initialize AI clients
-        let ollama_primary_url = env::var("OLLAMA_PRIMARY_URL").unwrap_or_else(|_| {
-            env::var("OLLAMA_URL").unwrap_or_else(|_| "http://localhost:11434".to_string())
-        });
-        let ollama_fallback_url = env::var("OLLAMA_FALLBACK_URL").ok();
-        let ollama_primary_model = env::var("OLLAMA_PRIMARY_MODEL")
-            .or_else(|_| env::var("OLLAMA_MODEL"))
-            .unwrap_or_else(|_| "nemotron-3-nano:30b".to_string());
-        let ollama_fallback_model = env::var("OLLAMA_FALLBACK_MODEL").ok();
-
-        let ollama = OllamaClient::new(
-            ollama_primary_url,
-            ollama_fallback_url,
-            ollama_primary_model,
-            ollama_fallback_model,
-        );
+        let ollama = build_ollama_from_env();
 
         let openrouter = build_openrouter_from_env();
         let openrouter_allowed_models = parse_openrouter_allowed_models();
@@ -375,13 +361,29 @@ fn parse_openrouter_allowed_models() -> Vec<String> {
         .collect()
 }
 
+/// Build the `OllamaClient` from environment variables — the canonical
+/// `OLLAMA_*` wiring shared by the server (`AppState::default`) and the
+/// standalone binaries (which predate this helper and used to copy it).
+pub fn build_ollama_from_env() -> OllamaClient {
+    let primary_url = env::var("OLLAMA_PRIMARY_URL").unwrap_or_else(|_| {
+        env::var("OLLAMA_URL").unwrap_or_else(|_| "http://localhost:11434".to_string())
+    });
+    let fallback_url = env::var("OLLAMA_FALLBACK_URL").ok();
+    let primary_model = env::var("OLLAMA_PRIMARY_MODEL")
+        .or_else(|_| env::var("OLLAMA_MODEL"))
+        .unwrap_or_else(|_| "nemotron-3-nano:30b".to_string());
+    let fallback_model = env::var("OLLAMA_FALLBACK_MODEL").ok();
+
+    OllamaClient::new(primary_url, fallback_url, primary_model, fallback_model)
+}
+
 /// Build a `LlamaCppClient` from environment variables. Returns `None` when
 /// `LLAMA_SWAP_URL` is unset. The client is constructed unconditionally
 /// when the URL is set (so it's available even under `LLM_BACKEND=ollama`
 /// for ad-hoc tooling), but the agentic / chat paths only route through it
 /// when `LLM_BACKEND=llamacpp`. Slot ids default to the names the bundled
 /// `llama-swap/config.yaml` uses — `chat` / `vision` / `embed`.
-fn build_llamacpp_from_env() -> Option<Arc<LlamaCppClient>> {
+pub fn build_llamacpp_from_env() -> Option<Arc<LlamaCppClient>> {
     let base_url = env::var("LLAMA_SWAP_URL").ok()?;
     let primary_model = env::var("LLAMA_SWAP_PRIMARY_MODEL").ok();
     let mut client = LlamaCppClient::new(Some(base_url), primary_model);

tts_pronunciations.example.json

@@ -0,0 +1,13 @@
+{
+  "Worcester": "Wuster",
+  "Spokane": "Spo can",
+  "wsl": "W S L",
+  "sql": "sequel",
+  "api": "A P I",
+  "US": "U S",
+  "Dr.": "Doctor",
+  "St.": "Saint",
+  "blvd": "boulevard",
+  "vs.": "versus",
+  "etc.": "et cetera"
+}