ImageApi/CLAUDE.md

# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

An Actix-web REST API for serving images and videos from a filesystem with automatic thumbnail generation, EXIF extraction, tag organization, and a memories feature for browsing photos by date. Uses SQLite/Diesel ORM for data persistence and ffmpeg for video processing.

## Development Commands

### Building & Running
```bash
# Build for development
cargo build

# Build for release (uses thin LTO optimization)
cargo build --release

# Run the server (requires .env file with DATABASE_URL, BASE_PATH, THUMBNAILS, VIDEO_PATH, BIND_URL, SECRET_KEY)
cargo run

# Run with specific log level
RUST_LOG=debug cargo run
```

### Testing
```bash
# Run all tests (requires BASE_PATH in .env)
cargo test

# Run specific test
cargo test test_name

# Run tests with output
cargo test -- --nocapture
```

### Database Migrations
```bash
# Install diesel CLI (one-time setup)
cargo install diesel_cli --no-default-features --features sqlite

# Create new migration
diesel migration generate migration_name

# Run migrations (also runs automatically on app startup)
diesel migration run

# Revert last migration
diesel migration revert

# Regenerate schema.rs after manual migration changes
diesel print-schema > src/database/schema.rs
```

### Code Quality
```bash
# Format code
cargo fmt

# Run clippy linter
cargo clippy

# Fix automatically fixable issues
cargo fix
```

### Utility Binaries
```bash
# Two-phase cleanup: resolve missing files and validate file types
cargo run --bin cleanup_files -- --base-path /path/to/media --database-url ./database.db
```

## Architecture Overview

### Core Components

**Layered Architecture:**
- **HTTP Layer** (`main.rs`): Route handlers for images, videos, metadata, tags, favorites, memories
- **Auth Layer** (`auth.rs`): JWT token validation, Claims extraction via FromRequest trait
- **Service Layer** (`files.rs`, `exif.rs`, `memories.rs`): Business logic for file operations and EXIF extraction
- **DAO Layer** (`database/mod.rs`): Trait-based data access (ExifDao, UserDao, FavoriteDao, TagDao)
- **Database Layer**: Diesel ORM with SQLite, schema in `database/schema.rs`

**Async Actor System (Actix):**
- `StreamActor`: Manages ffmpeg video processing lifecycle
- `VideoPlaylistManager`: Scans directories and queues videos
- `PlaylistGenerator`: Creates HLS playlists for video streaming

### Database Schema & Patterns

**Tables:**
- `users`: Authentication (id, username, password_hash)
- `favorites`: User-specific favorites (userid, path)
- `tags`: Custom labels with timestamps
- `tagged_photo`: Many-to-many photo-tag relationships
- `image_exif`: Rich metadata (file_path + 16 EXIF fields: camera, GPS, dates, exposure settings)

**DAO Pattern:**
All database access goes through trait-based DAOs (e.g., `ExifDao`, `SqliteExifDao`). Connection pooling uses `Arc<Mutex<SqliteConnection>>`. All DB operations are traced with OpenTelemetry in release builds.

**Key DAO Methods:**
- `store_exif()`, `get_exif()`, `get_exif_batch()`: EXIF CRUD operations
- `query_by_exif()`: Complex filtering by camera, GPS bounds, date ranges
- Batch operations minimize DB hits during file watching

### Multi-library data model

ImageApi supports more than one library (a library = a `(name, root_path)`
row in the `libraries` table that maps to a mounted directory tree). The
same bytes may exist under more than one library — typical case is an
"active" library plus an "archive" library that ingests files as they age
out — and the data model is designed so that derived data follows the
**bytes**, not the path, while user-managed data does the same.

**The principle.** A photo's identity is its `content_hash` (blake3, see
`src/content_hash.rs`). Anything we compute from or attach to a photo is
keyed on that hash so it survives:
- the same file appearing in a second library (backup / archive / mirror),
- the file moving between libraries (recent → archive handoff),
- the file moving within a library (re-organized rel_path),
- intra-library duplicates (same bytes at two paths).

**Table classification.** Three categories drive the keying decision:

| Category | Key | Rationale | Tables |
|---|---|---|---|
| Intrinsic to bytes | `content_hash` | Rerunning is wasted work (or LLM cost) | `face_detections` ✓, `image_exif` (target), `photo_insights` (target), `video_preview_clips` (target) |
| User intent about a photo | `content_hash` | "Tag this photo" means the bytes, not a path | `tagged_photo` (target), `favorites` (target) |
| Library administrative | `(library_id, rel_path)` | Tied to a specific filesystem location | `libraries`, `entity_photo_links`, the `rel_path` back-ref columns on hash-keyed tables |

✓ = already implemented this way. *(target)* = today still keyed on
`(library_id, rel_path)` and slated for migration. The migration adds a
nullable `content_hash` column, populates it from `image_exif` where
known, and read paths fall back to rel_path while the hash is null.

**Carrying a `rel_path` even when hash-keyed.** Hash-keyed tables retain
`(library_id, rel_path)` columns as a denormalized **back-reference**, not
as the key. This lets a single query answer "what is at this path right
now" without joining through `image_exif`, and supports the path-only
endpoints that predate the hash. `face_detections` is the reference
implementation: hash is the truth, path is a hint.

**Merge semantics on read.** When the same hash has rows under more than
one library:
- Set-valued data (tags, favorites, faces, entity links) → **union**.
- Scalar data (current insight, EXIF row, video preview clip) → earliest
  `generated_at` / `created_time` wins. The historical lib1 row beats a
  re-generated lib2 row, so the user's curated insight isn't shadowed by
  a re-run on archive ingest.

**Write attribution.** A new tag/favorite/insight created while viewing
under lib2 binds to the bytes, not to lib2 — so it shows up under lib1
too. This is by design, but it's the most surprising rule on first
encounter; clients should not assume tags are library-scoped.

**Hash-less rows (transitional state).** During and immediately after a
new mount, `image_exif.content_hash` is being populated by
`backfill_unhashed_backlog` (capped per tick). Rules during this window:
- Writes: if the hash is known, write hash-keyed. If not, write
  `(library_id, rel_path)`-keyed and let the reconciliation job collapse
  duplicates once the hash lands.
- Reads: prefer hash key, fall back to `(library_id, rel_path)`.
- Reconciliation: a one-shot pass after every backfill tick collapses
  rows that now share a hash, applying the merge semantics above.
  Idempotent — safe to re-run.

**Library handoff (recent → archive).** When a file moves between
libraries (e.g. operator moves `~/photos/2024/IMG.nef` to the archive
mount), the file watcher sees the disappearance under lib1 and the
appearance under lib2. Hash-keyed rows don't need migration; the
`(library_id, rel_path)` back-ref columns are updated to point to the new
location. Library administrative rows (`entity_photo_links`,
`(library_id, rel_path)` rows in `image_exif` for hash-less items) are
re-keyed by the move detector, which matches a disappearance to an
appearance by `content_hash` within a configurable window.

**Orphans (source deleted while a copy survives).** When the only
`image_exif` row for a hash is deleted (file removed from disk), the
hash-keyed derived rows survive **as long as another `image_exif` row
references the same hash**. If the last reference is gone, derived rows
are eligible for GC (deferred — the GC job runs on a slow schedule so
that a brief unmount or rename doesn't wipe history).

**Stats and counts.** When reporting "how many photos do you have," count
`DISTINCT content_hash` over `image_exif`, not row count. Faces stats
already does this (`FaceDao::stats` in `src/faces.rs`); other counters
should follow suit. Numerator and denominator must live in the same
domain — see the face-stats commentary below for the cautionary tale.

**Per-library scoping when the user asks for it.** A request scoped to
`?library=N` filters the `image_exif` view to that library, and the
hash-keyed derived data is joined through that view. The user sees only
photos that have a copy under lib N, but the derived data attached to
those photos is the merged hash-keyed view. This is the answer to "show
me archive photos with their original tags."

**Library availability and safety.** Libraries can be on network shares
or removable media; the file watcher must not interpret a temporary
unavailability as a mass-deletion event. Every tick begins with a
**presence probe** per library: the library is considered online iff
its `root_path` exists, is readable, and a top-level scan returns at
least one expected entry (or matches a recent file-count high-water
mark within a tolerance). The probe result gates which actions are safe
to run on that library this tick:

| Action | Requires online? |
|---|---|
| Quick / full scan ingest of new files | yes |
| EXIF / face / insight backlog drains | yes — but the work runs against any online library |
| Move-handoff detection (lib1 disappearance ↔ lib2 appearance match) | **both** libraries online |
| `(library_id, rel_path)` re-keying on detected move | **both** libraries online |
| Orphan GC of hash-keyed derived data | all libraries that have *ever* held the hash must be online and confirmed-clean for two consecutive ticks |
| Reads / serving | always allowed; falls back to whichever library is online |

A library that fails the probe enters a "stale" state: writes scoped to
it are paused, its rows are flagged stale (not deleted) in
`/libraries` status, and the watcher logs at `warn` once per
state-transition (not per tick). A library that recovers re-enters the
online set automatically; no operator action required for transient
outages. The intent is that pulling a USB drive, rebooting a NAS, or
losing a VPN never triggers a destructive code path — the worst case is
that derived-data work pauses until the share returns.

The same rule constrains the move-handoff matcher: a disappearance
under lib1 only counts as a "move" if there is a matching appearance
under another **online** library within the window. A bare
disappearance with no matching appearance is treated as
"unavailable-or-deleted, defer judgment" — it does not re-key any rows
and does not enqueue GC.

### File Processing Pipeline

**Thumbnail Generation:**
1. Startup scan: Rayon parallel walk of BASE_PATH
2. Creates 200x200 thumbnails in THUMBNAILS directory (mirrors source structure)
3. Videos: extracts frame at 3-second mark via ffmpeg
4. Images: uses `image` crate for JPEG/PNG processing
5. RAW formats (NEF/CR2/ARW/DNG/etc.): the `image` crate can't decode RAW
   pixel data, so the pipeline pulls an embedded JPEG preview instead. Fast
   path is `exif::read_jpeg_at_ifd` against IFD0 (PRIMARY) and IFD1
   (THUMBNAIL) — covers most older bodies and DNGs. Slow-path fallback shells
   out to **`exiftool`** for `PreviewImage` / `JpgFromRaw` / `OtherImage`,
   which reaches MakerNote / SubIFD-hosted previews kamadak-exif can't see
   (e.g. Nikon's `PreviewIFD`, where modern Nikon bodies store the full-res
   review JPEG). All candidates are pooled and the largest valid JPEG wins.
   See `src/exif.rs::extract_embedded_jpeg_preview`.

**File Watching:**
Runs in background thread with two-tier strategy:
- **Quick scan** (default 60s): Recently modified files only
- **Full scan** (default 3600s): Comprehensive directory check
- Batch queries EXIF DB to detect new files
- Configurable via `WATCH_QUICK_INTERVAL_SECONDS` and `WATCH_FULL_INTERVAL_SECONDS`

**EXIF Extraction:**
- Uses `kamadak-exif` crate
- Supports: JPEG, TIFF, RAW (NEF, CR2, CR3), HEIF/HEIC, PNG, WebP
- Extracts: camera make/model, lens, dimensions, GPS coordinates, focal length, aperture, shutter speed, ISO, date taken
- Triggered on upload and during file watching

**File Upload Behavior:**
If file exists, appends timestamp to filename (`photo_1735124234.jpg`) to preserve history without overwrites.

### Authentication Flow

**Login:**
1. POST `/login` with username/password
2. Verify with `bcrypt::verify()` against password_hash
3. Generate JWT with claims: `{ sub: user_id, exp: 5_days_from_now }`
4. Sign with HS256 using `SECRET_KEY` environment variable

**Authorization:**
All protected endpoints extract `Claims` via `FromRequest` trait implementation. Token passed as `Authorization: Bearer <token>` header.

### API Structure

**Key Endpoint Patterns:**

```rust
// Image serving & upload
GET  /image?path=...&size=...&format=...
POST /image (multipart file upload)

// Metadata & EXIF
GET /image/metadata?path=...

// Advanced search with filters
GET /photos?path=...&recursive=true&sort=DateTakenDesc&camera_make=Canon&gps_lat=...&gps_lon=...&gps_radius_km=10&date_from=...&date_to=...&tag_ids=1,2,3&media_type=Photo

// Video streaming (HLS)
POST /video/generate (creates .m3u8 playlist + .ts segments)
GET  /video/stream?path=... (serves playlist)

// Tags
GET    /image/tags/all
POST   /image/tags (add tag to file)
DELETE /image/tags (remove tag from file)
POST   /image/tags/batch (bulk tag updates)

// Memories (week-based grouping)
GET /memories?path=...&recursive=true

// AI Insights
POST /insights/generate              (non-agentic single-shot)
POST /insights/generate/agentic      (tool-calling loop; body: { file_path, backend?, model?, ... })
GET  /insights?path=...&library=...
GET  /insights/models                (local Ollama models + capabilities)
GET  /insights/openrouter/models     (curated OpenRouter allowlist)
POST /insights/rate                  (thumbs up/down for training data)

// Insight Chat Continuation
POST /insights/chat                  (single-turn reply, non-streaming)
POST /insights/chat/stream           (SSE: text / tool_call / tool_result / truncated / done)
GET  /insights/chat/history?path=... (rendered transcript with tool invocations)
POST /insights/chat/rewind           (truncate transcript at a rendered index)
```

**Request Types:**
- `FilesRequest`: Supports complex filtering (tags, EXIF fields, GPS radius, date ranges)
- `SortType`: Shuffle, NameAsc/Desc, TagCountAsc/Desc, DateTakenAsc/Desc

### Important Patterns

**Service Builder Pattern:**
Routes are registered via composable `ServiceBuilder` trait in `service.rs`. Allows modular feature addition.

**Path Validation:**
Always use `is_valid_full_path(&base_path, &requested_path, check_exists)` to prevent directory traversal attacks.

**File Type Detection:**
Centralized in `file_types.rs` with constants `IMAGE_EXTENSIONS` and `VIDEO_EXTENSIONS`. Provides both `Path` and `DirEntry` variants for performance.

**OpenTelemetry Tracing:**
All database operations and HTTP handlers wrapped in spans. In release builds, exports to OTLP endpoint via `OTLP_OTLS_ENDPOINT`. Debug builds use basic logger.

**Memory Exclusion:**
`PathExcluder` in `memories.rs` filters out directories from memories API via `EXCLUDED_DIRS` environment variable (comma-separated paths or substring patterns). The same excluder is applied to face-detection candidates (`face_watch::filter_excluded`) so junk directories like `@eaDir` / `.thumbnails` don't burn detect calls on Apollo.

### Face detection system

ImageApi owns the face data; Apollo (sibling repo) hosts the insightface inference service. Inference is triggered automatically by the file watcher and persisted into two tables:

- `persons(id, name UNIQUE COLLATE NOCASE, cover_face_id, entity_id, created_from_tag, notes, ...)` — operator-managed, name is the user-visible identity.
- `face_detections(id, library_id, content_hash, rel_path, bbox_*, embedding BLOB, confidence, source, person_id, status, model_version, ...)` — keyed on `content_hash` so a photo duplicated across libraries is detected once. Marker rows for `status IN ('no_faces','failed')` carry NULL bbox/embedding (CHECK constraint enforces this).

**Why content_hash and not (library_id, rel_path):** ties face data to the bytes, not the path. A backup mount that copies files from the primary library naturally inherits the existing detections without re-running inference. This is the reference implementation of the multi-library data model — see "Multi-library data model" above.

**File-watch hook** (`src/main.rs::process_new_files`): for each photo with a populated `content_hash`, check `FaceDao::already_scanned(hash)`; if not, send bytes (or embedded JPEG preview for RAW via `exif::extract_embedded_jpeg_preview`) to Apollo's `/api/internal/faces/detect`. K=`FACE_DETECT_CONCURRENCY` (default 8) parallel calls per scan tick; Apollo serializes them via its single-worker GPU pool. `face_watch.rs` is the Tokio orchestration layer.

**Per-tick backlog drain** (also `src/main.rs`): two passes that run on every watcher tick regardless of quick-vs-full scan:
- `backfill_unhashed_backlog` — populates `image_exif.content_hash` for photos that arrived before the hash field was retroactive. Capped by `FACE_HASH_BACKFILL_MAX_PER_TICK` (default 2000); errors don't burn the cap.
- `process_face_backlog` — runs detection on photos that have a hash but no `face_detections` row. Capped by `FACE_BACKLOG_MAX_PER_TICK` (default 64). Selected via a SQL anti-join (`FaceDao::list_unscanned_candidates`); videos and EXCLUDED_DIRS paths filtered out client-side via `face_watch::filter_excluded` so they never reach Apollo.

**Auto-bind on detection:** when a photo carries a tag whose name matches a `persons.name` (case-insensitive), the new face binds automatically iff cosine similarity to the person's existing-face mean is ≥ `FACE_AUTOBIND_MIN_COS` (default 0.4). Persons with no existing faces bind unconditionally and the new face becomes the cover.

**Manual face create** (`POST /image/faces`): crops the image to the user-supplied bbox, applies EXIF orientation via `exif::apply_orientation` (the `image` crate hands raw pre-rotation pixels — without this, manually-drawn bboxes never resolved a face on re-detection), pads to ~50% of bbox dims (RetinaFace anchor scales need ~50% face-fill at det_size=640), then calls Apollo's embed endpoint. A `force` flag lets the operator save a face the detector couldn't see (e.g. profile shots, occluded faces) — the row gets a zero-vector embedding so it's manually-bound only and won't participate in clustering.

**Rerun preserves manual rows** (`POST /image/faces/{id}/rerun`): only `source='auto'` rows are deleted before re-running detection. `already_scanned` returns true on ANY row, so a photo whose only faces are manually drawn never auto-redetects.

**Stats domain — content_hash, not file rows** (`FaceDao::stats` in `src/faces.rs`): `total_photos` counts `DISTINCT content_hash` over `image_exif` (filtered to image extensions, `content_hash IS NOT NULL`), and so do `scanned` / `with_faces` / `no_faces` / `failed` over `face_detections`. Numerator and denominator must live in the same domain — `face_detections` is keyed on content_hash, so the same JPEG present at two rel_paths or in two libraries scans once. Counting `image_exif` rows in the denominator inflated total by one per duplicate file and produced a permanent gap (e.g. 1101/1103 with nothing actually pending). Hash-less rows are excluded from total_photos while they sit in the `backfill_unhashed_backlog` queue; otherwise the bar pins below 100% for the duration of that backfill even though those rows aren't pending detection yet — they're pending hashing.

Module map:
- `src/faces.rs` — `FaceDao` trait + `SqliteFaceDao` impl, route handlers for `/faces/*`, `/image/faces/*`, `/persons/*`. Mirror of `tags.rs` layout.
- `src/face_watch.rs` — Tokio orchestration for the file-watch detect pass; `filter_excluded` (PathExcluder + image-extension filter), `read_image_bytes_for_detect` (RAW preview fallback).
- `src/ai/face_client.rs` — HTTP client for Apollo's inference. Configured by `APOLLO_FACE_API_BASE_URL`, falls back to `APOLLO_API_BASE_URL`. Both unset → feature disabled, file-watch hook is a no-op.
- `migrations/2026-04-29-000000_add_faces/` — schema.

### Startup Sequence

1. Load `.env` file
2. Run embedded Diesel migrations
3. Spawn file watcher thread
4. Create initial thumbnails (parallel scan)
5. Generate video GIF thumbnails
6. Initialize AppState with Actix actors
7. Set up Prometheus metrics (`imageserver_image_total`, `imageserver_video_total`)
8. Scan directory for videos and queue HLS processing
9. Start HTTP server on `BIND_URL` + localhost:8088

## Testing Patterns

Tests require `BASE_PATH` environment variable. Many integration tests create temporary directories and files.

When testing database code:
- Use in-memory SQLite: `DATABASE_URL=":memory:"`
- Run migrations in test setup
- Clean up with `DROP TABLE` or use `#[serial]` from `serial_test` crate if parallel tests conflict

## Common Gotchas

**EXIF Date Parsing:**
Multiple formats supported (EXIF DateTime, ISO8601, Unix timestamp). Fallback chain attempts multiple parsers.

**Video Processing:**
ffmpeg processes run asynchronously via actors. Use `StreamActor` to track completion. HLS segments written to `VIDEO_PATH`.

**File Extensions:**
Extension detection is case-insensitive. Use `file_types.rs` helpers rather than manual string matching.

**Migration Workflow:**
After creating a migration, manually edit the SQL, then regenerate `schema.rs` with `diesel print-schema`. Migrations auto-run on startup via `embedded_migrations!()` macro.

**Path Absolutization:**
Use `path-absolutize` crate's `.absolutize()` method when converting user-provided paths to ensure they're within `BASE_PATH`.

## Required Environment Variables

```bash
DATABASE_URL=./database.db              # SQLite database path
BASE_PATH=/path/to/media               # Root media directory
THUMBNAILS=/path/to/thumbnails         # Thumbnail storage
VIDEO_PATH=/path/to/video/hls          # HLS playlist output
GIFS_DIRECTORY=/path/to/gifs           # Video GIF thumbnails
BIND_URL=0.0.0.0:8080                  # Server binding
CORS_ALLOWED_ORIGINS=http://localhost:3000
SECRET_KEY=your-secret-key-here        # JWT signing secret
RUST_LOG=info                          # Log level
EXCLUDED_DIRS=/private,/archive        # Comma-separated paths to exclude from memories
```

Optional:
```bash
WATCH_QUICK_INTERVAL_SECONDS=60        # Quick scan interval
WATCH_FULL_INTERVAL_SECONDS=3600       # Full scan interval
OTLP_OTLS_ENDPOINT=http://...          # OpenTelemetry collector (release builds)

# AI Insights Configuration
OLLAMA_PRIMARY_URL=http://desktop:11434        # Primary Ollama server (e.g., desktop)
OLLAMA_FALLBACK_URL=http://server:11434        # Fallback Ollama server (optional, always-on)
OLLAMA_PRIMARY_MODEL=nemotron-3-nano:30b       # Model for primary server (default: nemotron-3-nano:30b)
OLLAMA_FALLBACK_MODEL=llama3.2:3b              # Model for fallback server (optional, uses primary if not set)
OLLAMA_REQUEST_TIMEOUT_SECONDS=120             # Per-request generation timeout (default 120). Increase for slow CPU-offloaded models.
SMS_API_URL=http://localhost:8000              # SMS message API endpoint (default: localhost:8000)
SMS_API_TOKEN=your-api-token                   # SMS API authentication token (optional)

# Apollo Places integration (optional). When set, photo-insight enrichment
# folds the user's personal place name (Home, Work, Cabin, ...) into the
# location string fed to the LLM, and the agentic loop gains a
# `get_personal_place_at` tool. Unset = legacy Nominatim-only path.
APOLLO_API_BASE_URL=http://apollo.lan:8000     # Base URL of the sibling Apollo backend

# Face inference (optional). Apollo also hosts the insightface inference
# service; ImageApi calls it from the file-watch hook (Phase 3) and from
# the manual face-create endpoint. Falls back to APOLLO_API_BASE_URL when
# unset (typical single-Apollo deploy). Both unset = feature disabled.
APOLLO_FACE_API_BASE_URL=http://apollo.lan:8000 # Override if face service runs separately
FACE_AUTOBIND_MIN_COS=0.4                       # Phase 3: cosine-sim floor for tag-name auto-bind
FACE_DETECT_CONCURRENCY=8                       # Phase 3: per-scan-tick parallel detect calls
FACE_DETECT_TIMEOUT_SEC=60                      # reqwest client timeout (CPU inference can be slow)

# OpenRouter (Hybrid Backend) - keeps embeddings + vision local, routes chat to OpenRouter
OPENROUTER_API_KEY=sk-or-...                   # Required to enable hybrid backend
OPENROUTER_DEFAULT_MODEL=anthropic/claude-sonnet-4   # Used when client doesn't pick a model
OPENROUTER_ALLOWED_MODELS=openai/gpt-4o-mini,anthropic/claude-haiku-4-5,google/gemini-2.5-flash
                                                # Curated allowlist exposed to clients via
                                                # GET /insights/openrouter/models. Empty = no picker.
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1     # Override base URL (optional)
OPENROUTER_EMBEDDING_MODEL=openai/text-embedding-3-small  # Optional, embeddings stay local today
OPENROUTER_HTTP_REFERER=https://your-site.example    # Optional attribution header
OPENROUTER_APP_TITLE=ImageApi                  # Optional attribution header

# Insight Chat Continuation
AGENTIC_CHAT_MAX_ITERATIONS=6                  # Cap on tool-calling iterations per chat turn (default 6)
```

**AI Insights Fallback Behavior:**
- Primary server is tried first with its configured model (5-second connection timeout)
- On connection failure, automatically falls back to secondary server with its model (if configured)
- If `OLLAMA_FALLBACK_MODEL` not set, uses same model as primary server on fallback
- Total request timeout is 120 seconds to accommodate slow LLM inference
- Logs indicate which server and model was used (info level) and failover attempts (warn level)
- Backwards compatible: `OLLAMA_URL` and `OLLAMA_MODEL` still supported as fallbacks

**Model Discovery:**
The `OllamaClient` provides methods to query available models:
- `OllamaClient::list_models(url)` - Returns list of all models on a server
- `OllamaClient::is_model_available(url, model_name)` - Checks if a specific model exists

This allows runtime verification of model availability before generating insights.

**Hybrid Backend (OpenRouter):**
- Per-request opt-in via `backend=hybrid` on `POST /insights/generate/agentic`.
- Local Ollama still describes the image (vision); the description is inlined
  into the chat prompt and the agentic loop runs on OpenRouter.
- `request.model` (if provided) overrides `OPENROUTER_DEFAULT_MODEL` for that
  call. The mobile picker reads from `OPENROUTER_ALLOWED_MODELS`.
- No live capability precheck — the operator-curated allowlist is trusted.
  A bad model id surfaces as a chat-call error.
- `GET /insights/openrouter/models` returns `{ models, default_model, configured }`
  for client picker UIs.

**Insight Chat Continuation:**

After an agentic insight is generated, the full `Vec<ChatMessage>` transcript is
stored in `photo_insights.training_messages` and can be continued via the
chat endpoints. The `PhotoInsightResponse.has_training_messages` flag tells
clients whether chat is available for a given insight.

- `POST /insights/chat` runs one turn of the agentic loop against the replayed
  history. Body: `{ file_path, library?, user_message, model?, backend?, num_ctx?,
  temperature?, top_p?, top_k?, min_p?, max_iterations?, amend? }`.
- `POST /insights/chat/stream` is the SSE variant — same request body, response
  is `text/event-stream` with events: `iteration_start`, `text` (delta), `tool_call`,
  `tool_result`, `truncated`, `done`, plus a server-emitted `error_message` on
  failure. Preferred by the mobile client for live tool-chip updates.
- `GET /insights/chat/history?path=...&library=...` returns the rendered
  transcript. Each assistant message carries a `tools: [{name, arguments, result,
  result_truncated?}]` array with the tool invocations that led up to it. Tool
  results over 2000 chars are truncated with `result_truncated: true`.
- `POST /insights/chat/rewind` truncates the transcript at a given rendered
  index (drops that message + any tool-call scaffolding that preceded it + all
  later turns). Index 0 is protected. Used for "try again from here" flows.

Backend routing rules (matches agentic-insight generation):
- Stored `backend` on the insight row is authoritative by default.
- `request.backend` may override per-turn. `local -> hybrid` is rejected in
  v1 (would require on-the-fly visual-description rewrite); `hybrid -> local`
  replays verbatim since the description is already inlined as text.
- `request.model` overrides the chat model (an Ollama id in local mode, an
  OpenRouter id in hybrid mode).

Persistence:
- Append mode (default): re-serialize the full history and `UPDATE` the same
  row's `training_messages`.
- Amend mode (`amend: true`): regenerate the title, insert a new insight row
  via `store_insight` (auto-flips prior rows' `is_current=false`). Response
  surfaces the new row's id as `amended_insight_id`.

Per-`(library_id, file_path)` async mutex (`AppState.insight_chat.chat_locks`)
serialises concurrent turns on the same insight so the JSON blob doesn't race.

Context management is a soft bound: if the serialized history exceeds
`num_ctx - 2048` tokens (cheap 4-byte/token heuristic), the oldest
assistant-tool_call + tool_result pairs are dropped until under budget. The
initial user message (with any images) and system prompt are always preserved.
The `truncated` event / flag is surfaced to the client when a drop occurred.

Configurable env:
- `AGENTIC_CHAT_MAX_ITERATIONS` — cap on tool-calling iterations per turn
  (default 6). Per-request `max_iterations` is clamped to this cap.

**Apollo Places integration (optional):**

The sibling Apollo project (personal location-history viewer) owns
user-defined Places: `name + lat/lon + radius_m + description (+ optional
category)`. When `APOLLO_API_BASE_URL` is set, ImageApi queries
`/api/places/contains?lat=&lon=` to enrich the LLM prompt's location
string. See `src/ai/apollo_client.rs` and `src/ai/insight_generator.rs`:

- **Auto-enrichment** (always on when configured): the per-photo location
  resolver folds the most-specific containing Place ("Home — near
  Cambridge, MA" or "Home (My house in Cambridge) — near Cambridge, MA"
  when a description is set) into the location field of `combine_contexts`.
  Smallest-radius wins — Apollo sorts server-side, this code takes `[0]`.
- **Agentic tool** `get_personal_place_at(latitude, longitude)`: registered
  alongside `reverse_geocode` only when `apollo_enabled()` returns true.
  Returns "- Name [category]: description (radius N m)" lines, smallest
  radius first. The tool is **deliberately narrow** — no enumerate-all
  variant; auto-enrichment covers the photo-context path and the agentic
  tool covers ad-hoc lat/lon questions in chat continuation.

Failure modes degrade silently to the legacy Nominatim path: 5 s timeout,
errors logged at `warn`, empty results returned. Apollo's routes are
unauthenticated (single-user, LAN-trust); add JWT auth here + on Apollo's
side if exposing beyond a trusted network.

## Dependencies of Note

### Rust crates

- **actix-web**: HTTP framework
- **diesel**: ORM for SQLite
- **jsonwebtoken**: JWT implementation
- **kamadak-exif**: EXIF parsing
- **image**: Thumbnail generation
- **walkdir**: Directory traversal
- **rayon**: Parallel processing
- **opentelemetry**: Distributed tracing
- **bcrypt**: Password hashing
- **infer**: Magic number file type detection

### External binaries (must be on `PATH`)

- **`ffmpeg`** — video thumbnail extraction (`StreamActor`, HLS pipeline) and
  the HEIF/HEIC/NEF/ARW thumbnail fallback in `generate_image_thumbnail_ffmpeg`.
  Required for any deploy that holds video or HEIF files.
- **`exiftool`** — optional but strongly recommended for RAW-heavy libraries.
  The thumbnail pipeline shells out to it as the slow-path fallback for
  embedded preview extraction (Nikon MakerNote `PreviewIFD`, Canon SubIFDs,
  etc. — anything kamadak-exif's IFD0/IFD1 readers can't reach). Without
  exiftool installed, RAWs whose preview lives outside IFD0/IFD1 will fall
  through to ffmpeg, which often produces black thumbnails. Install via
  package manager: `apt install libimage-exiftool-perl`,
  `brew install exiftool`, `winget install OliverBetz.ExifTool`, or
  `choco install exiftool`.