New optional SamplingOverride forwarded to llama-server as
chat_template_kwargs.enable_thinking (gates Qwen3-style reasoning
blocks). None leaves the template default; other backends ignore it.
Wired through the agentic-insight and chat-turn request bodies/handlers.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ALL-mode over-constrains NL queries — the model maps several query words to
tags and few photos carry every one, zeroing the candidate set. Switch to
ANY (a photo matches if it has any named tag); the semantic CLIP ranking
provides precision within that pool. Exclude tags still filter out.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When structured filters are present they're the constraint and CLIP only ranks
within the candidate set, so drop the global similarity threshold for that
case. Previously the 0.2 whole-library threshold ran BEFORE intersecting with
the filters, discarding filter-matching photos that scored just under it (e.g.
a 2022 beach photo at 0.18) — producing after_struct_filter=0 even when matches
existed. Plain semantic (no filters) keeps the user's threshold.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The CLIP encode failure reason was only ever returned in the HTTP response
body, never logged server-side, making 502s from /photos/search opaque. Log
the underlying cause — network error to the URL, or the Apollo HTTP status +
response body — so CLIP-service problems are diagnosable from the ImageApi log.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pin the NL->structured translation to a small, fast model that can stay
co-resident with CLIP (and the chat model) so it never evicts them on a tight
VRAM budget. Precedence: UNIFIED_SEARCH_MODEL env > client-selected model >
configured default. Logs the effective model (backend.model()) so model A/B
tests are visible. Documented in .env.example.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Log the translated query (semantic/tags/place/date/media + has_struct), the
tag-filter file count, candidate-row + allowed-hash counts, and the CLIP
considered/hits/after-filter counts. Pinpoints which stage drops results to
zero (over-extracted filter, tag path mismatch, Any/All over-constraint, or
CLIP threshold). info-level for now while debugging.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add an optional `model` query param to /photos/search/unified, passed into
resolve_backend's overrides. The client sends the user's currently-selected
local model so the translation step reuses an already-loaded model instead of
forcing a llama-swap eviction + cold start. Falls back to the configured
default when absent. Still local only (no hybrid).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Composes the two existing engines (Path A orchestration):
- Translate NL -> StructuredQuery via local LLM, respecting LLM_BACKEND
(resolve_backend(Local) -> ollama or llama-swap; no hybrid).
- Forward-geocode the place name into a gps circle.
- Structured filters (tags/EXIF/geo/date/media) build a candidate set of EXIF
rows; CLIP ranks within it, joined by content_hash. Degenerate cases match
existing behavior: semantic-only -> plain CLIP; filters-only -> date-sorted.
- Echoes the interpreted query (incl. resolved place) for editable client chips.
Refactor: extracted reusable cores from clip_search (score_photos, resolve_hits,
parse_library_scope, score_error_response) shared by both endpoints. Removed the
Phase 1 allow-until-wired attributes now that nl_query + geo are consumed.
fmt + clippy clean; 23 backend tests pass (7 geo, 12 nl_query, 4 unified).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Foundation for the /photos/search/unified endpoint (Phase 2). Two new,
fully unit-tested pieces, not yet wired into a route (allow-until-wired,
mirroring llm_client.rs):
- ai/nl_query.rs: translate a free-text query into a StructuredQuery via one
grounded LLM call. Two-stage — the model emits names/ISO dates, then a pure
resolve step maps tag names against the real vocab and converts dates to
unix seconds. Hallucinated (non-vocab) tags are surfaced in unmatched_tags
rather than silently used as hard filters — the anti-noise guard. 12 tests.
- geo::forward_geocode + bbox_to_circle: resolve a place name to a circle via
Nominatim /search, collapsing the bounding box to centroid + circumscribing
radius so "Portland" and "Italy" both map onto the existing gps circle
filter with no schema change. Radius is the max centroid-to-corner distance
(corners aren't equidistant on a sphere). 4 tests.
fmt + clippy clean; 19 new tests pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Nothing reaped reels before, so the on-disk cache and ledger grew
unbounded — each night's daily reel is a new ~4MB file + ledger row that's
stale within ~26h.
- Pre-gen self-prune: after recording a reel, prune_superseded keeps the
newest PREGEN_KEEP_PER_SCOPE (2) rows per (span, library) and unlinks the
superseded reels' mp4+sidecar. Caps the ledger/disk at ~spans×libraries×2.
- On-disk sweeper (spawn_reel_cache_sweeper): every 24h, removes reel mp4s
with no ledger row and no live job older than REEL_CACHE_MAX_AGE_DAYS (7) —
bounding the on-demand cache, which has no ledger row and otherwise grows
forever — plus crashed-render cruft (.mp4.tmp/.concat.txt/orphan sidecars).
Runs regardless of REEL_PREGEN_ENABLED; disable with REEL_CACHE_SWEEP_ENABLED=0.
- New DAO methods prune_superseded + all_cache_keys (with tests); env knobs
documented in .env.example.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Past the key-aware dedup, any mp4 already at the cache key was not
pre-generated by us (no matching ledger row) — typically an on-demand
fast-scripted reel sharing the key after the max_segments alignment.
Adopting it recorded a ledger row pointing at the fast reel, silently
defeating agentic pre-gen. Drop the adopt-existing-mp4 shortcut and
always produce_reel (atomic overwrite). Worst case is one redundant
re-render if a prior run crashed between render and ledger write.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
exists_fresh only matched (span, library, render_version, age), so a
cache-key change that doesn't bump RENDER_VERSION (e.g. the max_segments
alignment, or any future selection-logic tweak) left last night's ledger
row looking 'fresh' — the nightly run would skip and the orphaned reel
would persist. Dedup now compares the stored cache_key to the freshly
computed key (and confirms the mp4 exists), so a changed key forces a
regen within the freshness window. exists_fresh stays as the HTTP
endpoint's fast gate.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
pregen_one hardcoded max_segments: 24 while create_reel_handler defaults
to DEFAULT_MAX_SEGMENTS (40). Since the cache key encodes the raw
max_segments, the pre-generated reel's key never matched the client's
on-demand request, so POST /reels cache-hit an older max=40 reel and the
agentic pre-gen file was left orphaned. Align to DEFAULT_MAX_SEGMENTS (as
the plan specified) so the on-demand cache-hit path serves the pre-gen
reel. Content is unchanged — the actual beat count is duration-budgeted
either way; only the key descriptor differed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- pregen_one recorded media_count as planned.len() (beat count); record
the actual media item total (media.len(), photos + clips) in both the
cache-hit and freshly-rendered ledger paths. Drops the redundant
photo_count binding.
- Replace upsert_prefs's insert-then-catch-error-then-update dance with a
single atomic INSERT ... ON CONFLICT(id) DO UPDATE. Explicit id=1 makes
the conflict target deterministic; explicit column .set((...)) keeps
None -> NULL overwrite semantics so the row mirrors the latest request
exactly, and genuine insert errors surface instead of being swallowed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
1. Drop the unregistered prefs_dao/reel_dao web::Data extractors from
create_reel_handler / precomputed_reel_handler and read the DAOs off
AppState instead (consistent with the scheduler). Missing app_data
would have 500'd every POST /reels and /reels/precomputed at runtime.
2. Restore the dropped 'return' in the cache-hit branch — without it a
cache hit fell through, overwrote the Done job with Queued, and
re-ran the whole TTS+render pipeline on every request.
3. Make secs_until_next_run_hour minute/second-accurate so a batch that
finishes inside the run hour sleeps ~24h instead of busy-looping
(wake, re-run, sleep 0) for the rest of the hour. Tests updated.
4. Prune photo/user-bound tools (get_file_tags, get_faces_in_photo,
recall_facts_for_photo, recall_facts_for_entity) from the agentic
reel scripter's allow-list — they no-op/error with the empty
file/user context and only burn iterations.
5. Align AGENTIC_SYSTEM_PROMPT's advertised tool list with the actual
(pruned) allow-list.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Implement end-to-end nightly pre-generation of memory reels with agentic
scripting that grounds narration in calendar, location, messages, and RAG.
Sections A-E from the plan:
A. Extract produce_reel pipeline core from run_reel_job with
ScripterMode::Fast/Agentic and progress callbacks.
B. Agentic scripter: factor run_readonly_tool_loop from the insight
generator, build read-only tool gate, prompt builder with GPS, and
generate_script_agentic with fallback to fast path.
C. Precomputed reels ledger (SQLite table + DAO), GET /reels/precomputed
handler with validity gate, GET /reels/by-key/{key}/video streaming,
and normalize_library_key helper.
D. Nightly scheduler: spawn_pregen_scheduler with configurable hour,
run_pregen_batch (day/week/month spans), pregen_one with dedup and
disk-check, secs_until_next_run_hour time math.
E. user_ai_prefs passive mirror table + DAO for param capture in
create_reel_handler and replay in the scheduler.
Also fixes resolve_library_param signature to take &[Library] and adds
resolve_library_param_state wrapper for AppState callers.
New files: migrations/2026-06-13-000000_add_precomputed_reels/,
migrations/2026-06-13-000010_add_user_ai_prefs/,
src/database/precomputed_reel_dao.rs,
src/database/user_ai_prefs_dao.rs
A clip beat capped playback at CLIP_SECONDS and filled the rest of the
narration with a tpad freeze-frame, so a clip stopped dead on its last
frame for a second or two before the transition — a glitchy pause that
stills don't have. Extract clip_beat_plan: the clip now plays for as
much of its beat as the source footage covers, and we freeze only when
the source is genuinely shorter than the narration. Bump RENDER_VERSION.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
0.08s read as too abrupt; 0.12s keeps the burst clearly snappier than the
0.35s held-shot fade without jarring. Bumps RENDER_VERSION.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Videos in a span now appear as clip beats: the first few seconds of the
video (capped at CLIP_SECONDS=5, and to the source length) filled to the
portrait canvas like photos, with its live audio ducked under the
narration (amix at 0.35). If the narration outlasts the clip, the last
frame is held (tpad); clips with no audio track just play under narration.
Selection splits the beat budget between photo beats and clip beats —
clips get up to half (≥1 when present), photos the rest — then merges
both back into chronological order. SegmentMedia gains a Clip variant;
beats carry `media` (photos or one clip) and the cache key tags P/C so a
path used as a still vs a clip differ.
Also drops the burst fade from 0.15s to 0.08s so a quick burst reads
clearly differently from a held shot. Bumps RENDER_VERSION.
The clip filtergraph (fill + duck-mix + last-frame hold) is unit-tested
but, like the rest of the ffmpeg path, wants a real render check on the
GPU host.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Restructures a reel around beats — one narration line over one or more
photos — instead of one line per photo. A single-photo beat is a held
shot; a multi-photo beat is a quick burst that flashes through several
moments of an event while the line is read. So a week/month reel can show
everything it spans without a narrated (and timed) segment per photo.
Selection (selector.rs):
- Duration budget: cap the number of narrated beats to ~REEL_TARGET_SECONDS
(default 90, env-tunable) so week/month reels don't run minutes long.
- Event clustering by time gap; when there are more events than the beat
budget, adjacent events merge so the whole span stays covered. Each beat
bursts up to MAX_BURST_PHOTOS (an even spread), so a 40-shot dinner
contributes a handful of quick frames, not forty narrated seconds.
Render (render.rs): a beat renders its photos as a concat of per-photo
fills (blurred-bg portrait, fps-before-fade) under one muxed narration;
burst photos get a snappier fade. beat_durations splits the narration
across the photos, stretching only if a long burst would flash too fast.
Adds high-level info logs across the steps (request → script → per-beat
narrate/render → join → done with elapsed) for visibility. Bumps
RENDER_VERSION to re-render cached reels.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The fade looked steppy/low-frame-rate because the filtergraph normalized
fps AFTER the fade filters: the brightness ramp was sampled at the looped
still's coarse input cadence, then duplicated up to 30fps. Move fps ahead
of the fades, pin the still's input framerate (-framerate), and force CFR
output (-r) so the dip ramps across a full 30 frames and plays steadily.
Ease narration expressiveness from 0.7 to 0.6 (still tunable via
REEL_TTS_EXAGGERATION). Bump RENDER_VERSION so existing reels re-render.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Fixes the "image is tiny" problem: a 1920x1080 landscape reel letterboxes
to a ~25%-height band on a portrait phone. Switch to a portrait 1080x1920
canvas and fill it per photo with a blurred, zoomed copy of the image
behind the sharp fitted photo — so the frame is always full regardless of
the photo's orientation, with no black bars and no cropping of the subject.
Add a quick 0.35s fade in/out baked into each segment so concatenated
photos dip smoothly instead of hard-cutting (fade-out lands in the
narration's silent tail, so speech isn't clipped). Drop the unused
Ken Burns branch — motion can return deliberately later.
Warm up the narration a touch: thread Chatterbox's `exaggeration` through
synthesize_serialized and default reels to 0.7 (tunable via
REEL_TTS_EXAGGERATION). Bump RENDER_VERSION so existing landscape reels
re-render.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The concat stage wrote to <key>.mp4.tmp (for an atomic publish-rename),
but ffmpeg infers the muxer from the output extension and can't map
.tmp to a format — "Unable to choose an output format". Force the mp4
muxer explicitly so the temp extension is irrelevant. Segment render,
NVENC, TTS, and scripting were already working end-to-end; this was the
only failure, at the final join.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New POST /reels + GET /reels/{id} (+ /video) build an MP4 slideshow of a
memory span (day/week/month), narrated by the LLM in a cloned voice.
Pipeline (src/reels/): a selector resolves which photos + reel metadata,
the scripter writes one narration line per photo via a single LLM call
(reusing each photo's cached insight as context — no fresh vision calls,
so reel generation stays off the GPU's vision slot), each line is
synthesized to speech, and the renderer assembles stills + narration via
ffmpeg. Jobs run in the background (mirroring the TTS speech-job
registry) since a reel takes minutes; the finished MP4 is cached on disk
keyed by the selection so a repeat request is instant.
The segment model is media-typed (Photo today) so a video-clip segment
(phase 2) and a nightly pre-render (phase 3) slot in without reworking
the pipeline. Ken Burns motion is implemented but defaulted off pending a
visual check on the GPU box.
Supporting changes:
- memories: extract gather_memory_items() so the reel selector reuses the
exact window/exclusion/tz/sort logic behind /memories.
- ai::tts: add synthesize_serialized() so reel narration honors the same
single-GPU permit + write lease as user TTS requests.
- video::ffmpeg: make get_duration_seconds() pub for narration timing.
- AppState: reels_path (REELS_DIRECTORY, defaults beside preview clips).
Pure logic (cache key, script parsing, ffmpeg arg/filter construction,
even sampling, segment timing) is unit-tested (26 tests). The runtime
path (ffmpeg render, TTS, LLM) needs a real run on the GPU host to verify
end-to-end — not exercisable in CI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Clones that don't start at 0:00 are tagged with where the reference
window begins (grandma-at1m32s-30s), so voices cloned from different
sections of the same source are distinguishable in the voice list.
Zero-start names keep the existing -30s form.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both voice creation endpoints (upload + from-library) now accept optional
start_seconds/duration_seconds, threaded to ffmpeg as -ss/-t, so the
reference window can target clean speech anywhere in a long recording
instead of always the first N seconds. Duration is clamped to the
LLAMA_SWAP_TTS_REF_SECONDS cap and the voice-name tag reflects the
actual window length.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A JSON map (TTS_PRONUNCIATIONS_PATH, default tts_pronunciations.json)
rewrites mispronounced words — place names, initialisms, dotted
abbreviations — to phonetic spellings before synthesis, applied after
markdown cleanup in both /tts/speech paths. Whole-word smartcase
matching (lowercase keys match any casing, uppercase keys exact),
longest key wins, hot-reloaded on mtime change with last-good fallback
on parse errors. See tts_pronunciations.example.json.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Models wrap the title line despite the prompt — "**Title: A Day in the
Woods**", "## Title: ...", bold around just the label — which made
parse_title_body's bare "Title:" prefix match fall through to the
fallbacks and leak asterisks into the stored title.
strip_title_markdown trims bold/italic markers, heading hashes,
backticks, and quotes from both ends; applied to the label line, the
extracted title, both fallback paths, and generate_photo_title (which
previously stripped only quotes).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Trialing Qwen3-Embedding-0.6B (1024-dim, instruct-prefixed queries)
against nomic required code changes at every hardcoded seam; now it's a
config flip plus a reembed_embeddings run.
- EMBEDDING_DIM env (default 768) replaces every hardcoded dim check:
daily summary / calendar / search / location DAOs, Ollama batch
validation, reembed_embeddings
- entities gains the dim guard it never had — a wrong-dim vector
silently kills dedup/recall (cosine over mismatched lengths is 0),
so store None and warn instead
- embed_query / embed_document split with EMBED_QUERY_PREFIX /
EMBED_DOCUMENT_PREFIX (literal \n expanded): retrieval models treat
the two sides differently — nomic wants search_query:/search_document:,
Qwen3 wants Instruct:...\nQuery: on queries only. All query-side
call sites and all corpus writers now declare their side.
- document the contract in CLAUDE.md: change the model or any of these
vars → re-run reembed_embeddings or search is garbage
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The GPU lease keeps per-request reqwest budgets from burning behind a
cross-model swap, but the job-level INSIGHT_GENERATION_TIMEOUT_SECS
wall-clock started at spawn — an insight queued behind a running TTS
synthesis parked its first chat call on the lease and timed out
("timeout after 180s") before chatterbox even finished loading.
Acquire-and-drop an LLM read lease before starting the job clock in
both insight handlers: the wait for the GPU happens before the
timeout begins, mirroring the per-request lease semantics. Dropped
immediately — holding it across the generation would deadlock the
chat calls' own lease acquisitions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Queries embedded via llama-swap were searching corpora embedded via
Ollama (measured: spaces diverged). Introduce LocalLlm — the local
Ollama + llama-swap pair with LLM_BACKEND dispatch baked in — and route
all embedding writers through it; anything embedding via a concrete
client reintroduces the bug.
- search_rag: embed the model's query verbatim (no metadata boilerplate),
make date optional — no time-decay when omitted, so "when did X
happen?" queries rank purely by similarity across all time
- reembed_embeddings bin: re-embed summaries / calendar / search /
knowledge entities via the active backend, with old-new cosine report
per table and truncate-and-retry for inputs over the embed server's
physical batch size
- import_calendar, import_search_history: embed through LocalLlm
- search_messages / get_sms_messages: render sender → recipient so sent
messages are attributable to a conversation
- insight job failures: store the one-line anyhow context chain ({:#})
instead of the Debug dump the client was shown verbatim
- serialize env_dispatch tests behind a lock (parallel-runner flake)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
llama-swap runs chat/vision/Chatterbox as a mutually-exclusive set on
one GPU and HOLDS a request for a non-resident model until the resident
model drains, then swaps. That hold burned the holder's reqwest timeout
(measured: a queued TTS lost 77s behind one LLM turn; an LLM request
behind a synthesis waited the entire remaining synth), so concurrent
insight + read-aloud timed out instead of queueing.
ai::gpu adds a fair RwLock lease acquired before each request is sent,
so cross-model waits happen before the HTTP timeout starts: chat/vision
share the read lease, TTS synthesis and voice-library ops (which spin
Chatterbox up) take the write lease, and embeddings take none (the
embed slot is in llama-swap's always-resident group). Speech jobs now
flip queued->running only after acquiring the GPU, letting the client
anchor its poll deadline to that transition.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- DELETE /tts/voices/{name}: remove a cloned voice via the llama-swap
passthrough (upstream chatterbox-tts-api exposes DELETE /voices/{name}).
- POST/GET/DELETE /tts/speech/jobs: durable job flow for long syntheses —
dispatch returns 202 + job id, the synth queues on the GPU permit instead
of fast-failing 429, and clients poll for the result (kept ~10 min).
- GET /tts/voices now serves an in-memory cache so listing voices doesn't
make llama-swap spin up the TTS model (evicting the resident LLM);
invalidated on create/delete, ?refresh=1 forces an upstream re-query.
- Created voice names are tagged with LLAMA_SWAP_TTS_REF_SECONDS (e.g.
grandma-30s) so the library shows which ref length produced each clone.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The history-truncation budget assumed an 8192-token context whenever a
chat request omitted num_ctx, while the llama-swap chat slots serve
20k-131k. Replayed transcripts past ~6k tokens were silently gutted
every turn — losing conversation history and destroying llama.cpp
KV-cache prefix reuse (full SWA re-prefill per turn).
Default is now 32768 (real conversations top out around 16k), with
AGENTIC_CHAT_DEFAULT_NUM_CTX to override per deploy, floored at
headroom + 1024. Explicit per-request num_ctx still wins.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Keep `cargo clippy --tests` clean alongside the agentic-loop changes:
alias backfill's five-element setup() tuple as SetupFixture
(type_complexity) and build the single-library health map via
std::slice::from_ref instead of cloning (unnecessary clone-to-slice).
No behavior change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A request carrying persona_id but no system_prompt used to fall back to
the neutral default voice. Both agentic generation
(generate_agentic_insight_handler) and chat bootstrap now resolve the
persona's stored prompt from the persona store, with precedence:
explicit non-blank client system_prompt > persona store lookup >
existing default ("default" persona id behaves the same — used if the
store has a row, neutral default otherwise). Resolution happens at the
handler / bootstrap entry where the DAO is reachable; internals are
unchanged. resolve_bootstrap_system_prompt takes the resolved persona
prompt as a second argument, with precedence tests.
Also in insight_chat:
- Sync chat_turn no longer persists the synthetic "Please write your
final answer now without calling any more tools." user message pushed
on iteration exhaustion — extracted both streaming variants'
synthetic_idx pattern into push/remove_synthetic_final_prompt (the
remove is a defensive no-op on index drift) and applied it to all
three loops; round-trip test included.
- Strip leaked <think> blocks from the final content persisted as the
reply in chat_turn and both streaming AgenticLoopOutcomes (mid-stream
TextDeltas are untouched; the raw transcript keeps the block).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Agentic-loop fixes in the generator:
- New recall_facts_for_entity tool (always-on, like recall_entities):
fetches facts for one entity by id so the model can follow up on
entities surfaced by recall_entities that aren't photo-linked
(recall_facts_for_photo only covers linked entities). Mirrors that
tool's persona scoping (PersonaFilter::Single) and the persona's
reviewed_only_facts filter exactly, and renders in the same
"Entity: ... / - predicate object" style. Wired through execute_tool
and the trajectory summarizer.
- Generation now resolves gates persona-aware:
current_gate_opts_for_persona(images_inline, Some((user_id,
persona_id))) instead of the None-defaulting wrapper, so a persona's
allow_agent_corrections opens propose_correction during generation the
same way chat turns already did. The now-unused current_gate_opts
wrapper is removed.
- Strip leaked <think> blocks from the final assistant content before
parse_title_body / store_insight (raw training transcript keeps them).
- Honest truncation labels: get_sms_messages and get_location_history
said "Found N ..." while listing only the first K; found_header now
emits "Found N ... (showing first K):" when truncated, and the
summarizer still parses the count.
- Clamp days_radius in get_calendar_events and get_location_history to
1..=30, matching get_sms_messages.
- persona_system_prompt helper (persona store lookup, blank-prompt ->
None) for server-side persona resolution; callers land in the next
commit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ollama >=0.8 can stream tool_calls incrementally across NDJSON chunks;
chat_with_tools_stream did `tool_calls = Some(tcs)` per chunk, so only
the last chunk's calls survived assembly and earlier calls were silently
dropped. Append into the accumulator instead.
- ollama: append_streamed_tool_calls helper + tests covering two calls
arriving in separate chunks and the single-chunk batch case.
- llamacpp: the SSE delta assembly was already correct (per-index
BTreeMap, same-index argument fragments concatenate, distinct indexes
accumulate); extracted it into apply_tool_call_deltas /
finalize_tool_calls and added tests pinning that behavior.
- llm_client: new shared strip_think_blocks (moved from ollama's private
extract_final_answer, which now delegates) so the tool-calling final
content paths can reuse it; unit tests for tagged/plain/unclosed/empty
cases.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Expose GET /insights/history?path=... returning every generated version
of a photo's insight (current plus superseded), newest-first, backing the
mobile per-file insight history view.
- New get_insight_history_handler; reuses the existing get_insight_history
DAO method (removed its dead_code allow).
- impl From<PhotoInsight> for PhotoInsightResponse, collapsing the mapping
that was duplicated across the single-get and all-insights handlers.
- rate_insight_by_id DAO method + optional insight_id on RateInsightRequest
so previously generated versions can be approved/rejected (the path-based
rate only touches the current row).
- DAO tests for history ordering/scoping and id-targeted rating.
- cargo fmt normalized a multi-line assert in insight_chat.rs tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Chatterbox inserts a long pause — sometimes ~20s of silence — for each
blank line it sees, and insight text is markdown full of paragraph
breaks. clean_for_tts previously preserved paragraph structure
(\n{3,} -> \n\n), so every paragraph boundary still reached the model
as a double newline. Now any run of 2+ newlines, including
whitespace-only blank lines, collapses to a single newline so the
worst pause a break can cause is a normal line-break pause.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Chatterbox wrapper has no internal lock or cancellation, so concurrent
synth requests contend on the single GPU and abandoned (timed-out) jobs
cascade into stacked slowness. Gate synthesis behind a one-permit semaphore
and fast-fail concurrent requests with 429 instead of queueing.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Long insights are chunked + synthesized server-side and can run past the shared
180s chat/embedding client timeout, causing spurious timeouts. /tts/speech now
uses a per-request timeout from LLAMA_SWAP_TTS_REQUEST_TIMEOUT_SECONDS
(default 600), overriding the client default without affecting chat/embeddings.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Each /tts handler now opens an http.tts.* span via extract_context_from_request
+ global_tracer().start_with_context, sets Status::Ok / Status::error on every
outcome, and records useful attributes (model, format, voice_name, byte counts)
— matching the insight handlers. Prometheus request metrics were already
covered by the app-wide actix-web-prom middleware.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Chatterbox validates the reference clip by file extension and rejects formats
like .aac/.opus. Always transcode the reference (upload bytes and library
files alike) to mono 24 kHz WAV with ffmpeg before forwarding, so any source
format is accepted and the from-library audio/video paths are unified.
The reference length cap is now configurable via LLAMA_SWAP_TTS_REF_SECONDS
(default 30) — Chatterbox is zero-shot, so a clean ~10-20s clip is the sweet
spot. Drops the now-unused mime guesser.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
/tts/speech now normalizes input before synthesis: unwraps markdown
links/images to visible text, drops heading/list/blockquote/emphasis
markers and URLs, strips emoji (which non-turbo Chatterbox mispronounces
or skips), and collapses whitespace. Centralized in clean_for_tts so the
app, WebUI, and curl all get clean audio. Bracketed tags are deliberately
preserved for a future Turbo (paralinguistic) switch.
Adds optional exaggeration / cfg_weight / temperature to the request,
clamped to Chatterbox's documented ranges and forwarded on the speech
body. Unit tests cover markdown/emoji/URL stripping and tag preservation.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
LlamaCppClient gains text_to_speech (OpenAI /audio/speech), list_voices and
create_voice (voice library at the swap-root /upstream/<model>/voices
passthrough), plus a tts_model slot configured via LLAMA_SWAP_TTS_MODEL
(default "chatterbox").
New Claims-gated routes:
- POST /tts/speech -> { audio_base64, format } for data: URI playback
- GET /tts/voices -> voice library passthrough
- POST /tts/voices/upload -> clone a voice from an uploaded clip (multipart)
- POST /tts/voices/from-library -> clone from a library file (ffmpeg-extracts
audio from video; audio forwarded as-is)
Security: voice_name sanitized to [A-Za-z0-9_-] (it becomes an upstream
filename), 25 MB upload cap, library refs restricted to real audio/video,
path confined via is_valid_full_path. Adds is_audio_file + unit tests for the
sanitizer, mime guesser, and swap-root derivation.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
probe_video_stream_meta requested a bare `side_data_list` section in
-show_entries. On modern ffprobe that's the *frame* side-data section,
so ffprobe enumerated every frame to collect it — reading the entire
mdat. For non-faststart phone clips on the SMB mount this turned a
metadata probe into a full-file read: /video/generate took 10-32s per
open (0% CPU, time proportional to file size).
Switch to `stream_side_data_list`, which reads the Display Matrix
rotation from the stream header (moov) without touching frames. Codec,
frame rate, and rotation are unchanged; the existing rotation parser
already reads streams[0].side_data_list[].rotation. Fixes both the
open-path probe and the transcode actor's probe. Cold opens now return
near-instantly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
generate_video built the rel_path for its image_exif lookup by stripping
the library root from the absolute path, leaving backslashes on Windows
(Melissa\clip.mp4). file_scan stores rel_paths forward-slash and
get_exif_batch matches exactly with no normalization, so the lookup
missed and the handler re-hashed the entire video file on every request.
Extract rel_path_for_lookup and normalize separators with replace('\\',
'/'). Adds tests for Windows/Unix separators, file-at-root, leading
separator stripping, and the no-match fallback.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The truncation budget estimated message size by serializing the full
ChatMessage array, including the base64 image persisted in the first
user message. A 1024px JPEG is hundreds of KB of base64 characters —
8-19x the entire ~24KB text budget at the default num_ctx — and the
image lives in the protected prefix that's never dropped. The budget
check was therefore essentially always over, dropping all tool history
and firing the "trimmed context" banner on every turn for vision
backends that inline images.
estimate_bytes now strips image payloads before counting and charges a
flat IMAGE_TOKENS_EACH per image instead, so the budget reflects real
text token pressure. Adds a regression test covering a short
conversation with one large image.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Cameron Cordes