Models page, per-model prompts, copy job, SQLite + Omnilingual ASR

.env.example

@@ -44,3 +44,17 @@ JOB_TIMEOUT=30m
 # above the largest burst the pipeline will submit.
 MAX_TERMINAL_JOBS=200
 LOG_FORMAT=json
+
+# ---- Omnilingual ASR (optional) ----
+# Meta Omnilingual ASR 7B runs as a separate container. Enable with
+#   docker compose --profile omnilingual up -d
+# and set OMNILINGUAL_URL so the transcriber registers the `omnilingual-7b`
+# model. Leave it empty to keep the model hidden.
+#
+# VRAM: the LLM 7B model holds ~17 GiB on the same GPU as whisper (~3 GB per
+# worker). On a 24 GB RTX 3090 set WORKERS=1 while this is on, or switch to
+# OMNI_MODEL_CARD=omniASR_CTC_7B_v2 (~15 GiB, much faster, ignores language).
+# OMNILINGUAL_URL=http://omnilingual:8000
+# OMNILINGUAL_IMAGE=ghcr.io/bcc-code/omnilingual-asr:latest
+# OMNI_MODEL_CARD=omniASR_LLM_7B_v2
+# OMNI_BATCH_SIZE=4

.github/workflows/omnilingual.yml

@@ -0,0 +1,84 @@
+name: omnilingual
+
+# Builds the Omnilingual ASR sidecar image (omnilingual/) and pushes it to GHCR
+# so hosts can `docker compose --profile omnilingual pull`. Only runs when the
+# sidecar changes. Model weights are not baked in — they download into the
+# omnilingual-cache volume on first start.
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - omnilingual/**
+      - .github/workflows/omnilingual.yml
+  pull_request:
+    branches: [main]
+    paths:
+      - omnilingual/**
+      - .github/workflows/omnilingual.yml
+  workflow_dispatch:
+
+env:
+  REGISTRY: ghcr.io
+  IMAGE_NAME: ${{ github.repository_owner }}/omnilingual-asr
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+      - uses: astral-sh/setup-uv@v6
+      # Only the dependency-free region planner; the server needs a GPU + 7B model.
+      - run: uv run --no-project --with pytest pytest -q test_regions.py
+        working-directory: omnilingual
+
+  image:
+    runs-on: ubuntu-latest
+    needs: test
+    env:
+      CAN_PUSH: ${{ github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository }}
+    permissions:
+      contents: read
+      packages: write
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Free up disk space
+        # torch + CUDA wheels are several GB installed; ubuntu-latest has ~14 GB free.
+        run: |
+          sudo rm -rf \
+            /usr/local/lib/android \
+            /usr/share/dotnet \
+            /opt/ghc \
+            /usr/local/share/boost \
+            /opt/hostedtoolcache/CodeQL
+          sudo docker image prune -af
+          df -h /
+
+      - uses: docker/setup-buildx-action@v3
+
+      - name: Log in to GHCR
+        if: env.CAN_PUSH == 'true'
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+          tags: |
+            type=ref,event=branch
+            type=sha,prefix=sha-
+            type=raw,value=latest,enable={{is_default_branch}}
+
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: omnilingual
+          platforms: linux/amd64
+          push: ${{ github.event_name != 'pull_request' && env.CAN_PUSH == 'true' }}
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}

.gitignore

@@ -33,3 +33,11 @@ prompt.txt
 
 # Compiled binary from `make build`
 /transcriber
+
+# Python (omnilingual sidecar)
+.venv
+__pycache__
+.pytest_cache
+
+# Local SQLite database (+ WAL/SHM sidecars)
+transcriber.db*

DEPLOY.md

@@ -236,6 +236,7 @@ inspecting a bad transcription.
 | Mount                                          | Purpose                                                                                                                                              |
 | ---------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `models:/var/cache/transcriber`                | ggml whisper.cpp models. Survives restarts — first job downloads ~3 GB.                                                                               |
+| `data:/var/lib/transcriber`                    | SQLite database (`transcriber.db`) — per-model prompts edited on the UI's Models page.                                                               |
 | `${STORAGE_PATH}:${STORAGE_PATH}`              | Audio inputs (`path`) and transcript outputs (`output_path`). Mounted at the same path inside and out, because request paths resolve in the container. |
 | `./prompt.txt:/app/prompt.txt:ro` _(optional)_ | Default prompt file. Uncomment in `docker-compose.yml`. Without it, only requests carrying their own `prompt` get one.                                 |
 
@@ -251,6 +252,36 @@ compute-bound here, not memory-bound. Override `WHISPER_CPP_MODEL` /
 `NB_WHISPER_MODEL` / `WHISPER_VAD_MODEL` on the service to pin a model to
 a specific file on disk instead of letting the HF cache resolve it.
 
+## Omnilingual ASR (optional)
+
+`omnilingual-7b` (Meta Omnilingual ASR 7B) runs in its own container, built
+from `omnilingual/`. It is a Python/PyTorch service and is off by default.
+
+```sh
+# .env
+OMNILINGUAL_URL=http://omnilingual:8000
+WORKERS=1                # see VRAM below
+
+docker compose --profile omnilingual up -d
+docker compose logs -f omnilingual    # wait for "model ready"
+```
+
+- **VRAM.** The LLM 7B checkpoint holds ~17 GiB, on the same GPU as
+  whisper (~3 GB per worker). On a 24 GB RTX 3090, `WORKERS=2` plus the
+  sidecar overshoots and jobs fail with CUDA OOM. Use `WORKERS=1`, or
+  `OMNI_MODEL_CARD=omniASR_CTC_7B_v2` (~15 GiB, much faster, ignores
+  language).
+- **Driver.** The image uses PyPI torch 2.8.0, which bundles a CUDA 12.8
+  runtime. That needs a newer host driver than the whisper image's CUDA 12.6
+  (>= 560.28.03). If the sidecar fails at CUDA init, check `nvidia-smi`
+  against NVIDIA's CUDA 12.8 driver table.
+- **First start** downloads the multi-GB checkpoint into the
+  `omnilingual-cache` volume. The healthcheck allows 30 minutes for it.
+  Jobs submitted earlier fail fast with `503 model not loaded`.
+- **Output.** There are no word timestamps, so subtitles get one cue per
+  speech region (at most 30 s). There is no prompt support, and the text is
+  likely lowercase with little punctuation. See `omnilingual/README.md`.
+
 ## Pre-seeding models
 
 Worth doing even with internet access: the download currently happens

Dockerfile

@@ -52,7 +52,8 @@ ARG ARCH
 # the build depend on whatever the floating image tag resolved to.
 ENV GOTOOLCHAIN=local CGO_ENABLED=0 GOOS=linux
 WORKDIR /src
-COPY go.mod ./
+COPY go.mod go.sum ./
+RUN go mod download
 COPY cmd ./cmd
 COPY internal ./internal
 # internal/web/dist is gitignored, so it is absent (or stale) in the build
@@ -68,7 +69,7 @@ FROM --platform=linux/${ARCH} ${BASE_IMAGE}
 # Model cache: $XDG_CACHE_HOME/transcriber/hf/<repo>/<file>
 ENV XDG_CACHE_HOME=/var/cache \
     WHISPER_CPP_BIN=/usr/local/bin/whisper-cli
-RUN mkdir -p /var/cache/transcriber/hf
+RUN mkdir -p /var/cache/transcriber/hf /var/lib/transcriber
 
 COPY --from=go-build /out/transcriber /usr/local/bin/transcriber
 
@@ -79,4 +80,4 @@ EXPOSE 8888
 # Flags come from docker-compose.yml (driven by .env); these defaults apply
 # when running the image directly with `docker run`.
 ENTRYPOINT ["/usr/local/bin/transcriber"]
-CMD ["-port=8888", "-workers=2", "-default-model=whisper-cpp-large-v3", "-log-format=json"]
+CMD ["-port=8888", "-workers=2", "-default-model=whisper-cpp-large-v3", "-db=/var/lib/transcriber/transcriber.db", "-log-format=json"]

Makefile

@@ -1,4 +1,4 @@
-.PHONY: help dev dev-api dev-frontend frontend build run test clean
+.PHONY: help dev dev-api dev-frontend frontend build run test sqlc clean
 
 help:
 	@echo 'Targets:'
@@ -9,6 +9,7 @@ help:
 	@echo '  build         frontend + go build → ./transcriber (single binary)'
 	@echo '  run           build then run ./transcriber'
 	@echo '  test          go test -race ./...'
+	@echo '  sqlc          regenerate internal/db/dbq from internal/db/{migrations,queries}'
 	@echo '  clean         remove build artifacts'
 
 # Run both servers in parallel. `set -m` puts each background recipe in its
@@ -47,6 +48,9 @@ run: build
 test:
 	go test -race ./...
 
+sqlc:
+	sqlc generate
+
 clean:
 	rm -f transcriber
 	rm -rf frontend/.nuxt frontend/.output

README.md

@@ -57,6 +57,7 @@ Go code. Server settings come from flags; per-machine paths from env vars.
 | `-callback-workers`    | `2`          | webhook delivery goroutines                                                                                                                                                           |
 | `-default-model`       | `stub`       | adapter ID used when the request omits `model`                                                                                                                                        |
 | `-default-prompt-file` | `prompt.txt` | file whose contents are used as the prompt when the request omits one; missing file = no default. A non-empty `prompt` in the request fully overrides it                              |
+| `-db`                 | `transcriber.db` | SQLite file, created and migrated (embedded goose migrations) on startup. Holds per-model prompts set on the UI's Models page |
 | `-job-timeout`         | `30m`        | wall-clock cap per job; on expiry the worker cancels the subprocess and marks the job `FAILED` with `error: "timeout"`. Per-request `timeout_seconds` overrides this. `<= 0` disables |
 | `-max-terminal-jobs`   | `20`         | how many finished jobs (completed/failed/canceled) to retain in memory; `<= 0` disables the cap                                                                                       |
 | `-log-format`          | `text`       | `text` for human-readable output (dev), `json` for structured logs (prod). The Dockerfile sets `json`                                                                                 |
@@ -70,6 +71,7 @@ Go code. Server settings come from flags; per-machine paths from env vars.
 | `NB_WHISPER_MODEL`  | _(unset → fetched from HF)_     | local path override for the `nb-whisper-large` adapter. Unset = auto-download `NbAiLab/nb-whisper-large/ggml-model.bin` via `internal/hfcache`.     |
 | `WHISPER_VAD_MODEL` | _(unset → fetched from HF)_     | local path override for the Silero VAD model used to skip music/silence. Unset = auto-download `ggml-org/whisper-vad/ggml-silero-v5.1.2.bin`.       |
 | `XDG_CACHE_HOME`    | `~/.cache`                      | base for the HF cache (`<root>/transcriber/hf/<repo>/<file>`).                                                                                      |
+| `OMNILINGUAL_URL`   | _(unset)_                       | base URL of the Omnilingual ASR sidecar (`omnilingual/`). When set, the `omnilingual-7b` model is registered; unset = model hidden.                  |
 
 ## API
 
@@ -133,10 +135,18 @@ serialize to.
 
 - `GET  /transcription/jobs` — list all jobs
 - `DELETE /transcription/job/{id}` — cancel a queued or running job
-- `GET  /models` — list registered adapters
+- `GET  /models` — list registered adapters (with aliases and per-model prompt)
+- `PUT  /models/{id}/prompt` — set/clear a model's default prompt. Precedence: request `prompt` > model prompt > `-default-prompt-file`
 - `GET  /config` — server-side defaults (currently `default_prompt`)
 - `GET  /healthz`, `GET /readyz`
 
+## Database
+
+SQLite (pure-Go `modernc.org/sqlite`, so `CGO_ENABLED=0` still works).
+Migrations live in `internal/db/migrations/` and are embedded and applied by
+goose on startup. Queries live in `internal/db/queries/`; after editing either,
+run `make sqlc` to regenerate `internal/db/dbq/`. Jobs are still in-memory only.
+
 ## Adding a new backend
 
 1. Implement `transcriber.Transcriber` in `internal/transcriber/<name>/`.

cmd/transcriber/main.go

@@ -14,7 +14,9 @@ import (
 
 	"transcriber/internal/api"
 	"transcriber/internal/callback"
+	"transcriber/internal/db"
 	"transcriber/internal/jobs"
+	"transcriber/internal/modelprompts"
 	"transcriber/internal/transcriber"
 	"transcriber/internal/web"
 	"transcriber/internal/worker"
@@ -27,6 +29,7 @@ func main() {
 	defaultModel := flag.String("default-model", "stub", "model adapter ID to use when the request omits `model`")
 	defaultLanguage := flag.String("default-language", "", "ISO 639-1 language hint applied when the request omits `language` (empty or `auto` = let the model detect)")
 	defaultPromptFile := flag.String("default-prompt-file", "prompt.txt", "path to a file whose contents are used as the prompt when the request omits one (missing file = no default prompt)")
+	dbPath := flag.String("db", "transcriber.db", "SQLite database file (created and migrated on startup); holds per-model prompts edited from the UI")
 	maxTerminalJobs := flag.Int("max-terminal-jobs", 20, "how many finished jobs (completed/failed/canceled) to retain in memory; <= 0 disables the cap")
 	jobTimeout := flag.Duration("job-timeout", 30*time.Minute, "default wall-clock timeout per job; per-request timeout_seconds overrides this; <= 0 disables")
 	logFormat := flag.String("log-format", "text", "log handler: text (human-readable, dev) or json (structured, prod)")
@@ -63,6 +66,14 @@ func main() {
 		}
 	}
 
+	conn, err := db.Open(context.Background(), *dbPath)
+	if err != nil {
+		slog.Error("opening database", "path", *dbPath, "err", err)
+		os.Exit(1)
+	}
+	defer conn.Close()
+	modelPrompts := modelprompts.New(conn)
+
 	registry := buildRegistry(*defaultModel, *whisperNoGPU)
 	if _, ok := registry.Default(); !ok {
 		slog.Error("default model not registered", "id", *defaultModel)
@@ -96,7 +107,7 @@ func main() {
 	})
 	pool.Start(ctx)
 
-	srv := api.NewServer(store, queue, registry, defaultPrompt, normalizedDefaultLang)
+	srv := api.NewServer(store, queue, registry, modelPrompts, defaultPrompt, normalizedDefaultLang)
 	addr := fmt.Sprintf(":%d", *port)
 	httpSrv := &http.Server{
 		Addr:              addr,

cmd/transcriber/models.go

@@ -7,6 +7,7 @@ import (
 	"transcriber/internal/hfcache"
 	"transcriber/internal/transcriber"
 	"transcriber/internal/transcriber/chunked"
+	"transcriber/internal/transcriber/omnilingual"
 	"transcriber/internal/transcriber/stub"
 	"transcriber/internal/transcriber/whispercpp"
 )
@@ -59,6 +60,20 @@ func buildRegistry(defaultID string, whisperNoGPU bool) *transcriber.Registry {
 		chunked.Config{},
 	))
 
+	// Meta Omnilingual ASR runs as a Python sidecar (omnilingual/). Only
+	// registered when its URL is configured, so /models never advertises a
+	// model whose backend isn't running.
+	if url := os.Getenv("OMNILINGUAL_URL"); url != "" {
+		r.Register(chunked.New(
+			omnilingual.New(omnilingual.Config{
+				ID:          "omnilingual-7b",
+				DisplayName: "Meta Omnilingual ASR 7B",
+				BaseURL:     url,
+			}),
+			chunked.Config{},
+		))
+	}
+
 	// Sub-large sizes upgrade to large — matches the old runner's coercion behavior.
 	r.Alias("openai/whisper-large-v3", "whisper-cpp-large-v3")
 	r.Alias("openai/whisper-large-v2", "whisper-cpp-large-v3")

docker-compose.yml

@@ -38,10 +38,18 @@ services:
       - "-job-timeout=${JOB_TIMEOUT:-30m}"
       - "-max-terminal-jobs=${MAX_TERMINAL_JOBS:-200}"
       - "-log-format=${LOG_FORMAT:-json}"
+      - "-db=/var/lib/transcriber/transcriber.db"
+
+    environment:
+      # Set to http://omnilingual:8000 when the `omnilingual` profile is on.
+      # Empty = the omnilingual-7b model is not registered.
+      OMNILINGUAL_URL: ${OMNILINGUAL_URL:-}
 
     volumes:
       # ggml model cache — first job downloads ~3 GB, this keeps it.
       - models:/var/cache/transcriber
+      # SQLite database (per-model prompts edited from the UI).
+      - data:/var/lib/transcriber
       # Audio in / transcripts out. Job requests carry absolute paths that are
       # resolved *inside* the container, so host and container paths must match.
       # STORAGE_PATH must already exist on the host: Docker otherwise creates
@@ -65,5 +73,38 @@ services:
               count: all
               capabilities: [gpu]
 
+  # Meta Omnilingual ASR sidecar (omnilingual/). Opt-in:
+  #   docker compose --profile omnilingual up -d
+  # and set OMNILINGUAL_URL=http://omnilingual:8000 in .env. It holds ~17 GiB
+  # of VRAM on the same GPU as whisper. See DEPLOY.md before enabling it.
+  omnilingual:
+    profiles: ["omnilingual"]
+    image: ${OMNILINGUAL_IMAGE:-ghcr.io/bcc-code/omnilingual-asr:latest}
+    build:
+      context: ./omnilingual
+    restart: unless-stopped
+    environment:
+      OMNI_MODEL_CARD: ${OMNI_MODEL_CARD:-omniASR_LLM_7B_v2}
+      OMNI_BATCH_SIZE: ${OMNI_BATCH_SIZE:-4}
+    volumes:
+      # fairseq2 checkpoint cache — the 7B model is a multi-GB first download.
+      - omnilingual-cache:/cache
+    healthcheck:
+      test: ["CMD", "curl", "-fsS", "http://localhost:8000/healthz"]
+      interval: 30s
+      timeout: 5s
+      # First start downloads and loads the 7B checkpoint.
+      start_period: 30m
+      retries: 3
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+
 volumes:
   models:
+  data:
+  omnilingual-cache: