Refactor/consolidate backend services

[Ayan-josh-05]

Consolidate Backend Services into a Single Container

Summary

Collapses app, translation, field_mapping, ocr, and gateway from five separate containers into one (backend / backend-gpu), taking the stack from 9 containers down to 5.

None of the five services' own code changes — they keep running as independent FastAPI apps on their existing internal ports (8000/8001/8002/8010/8080), just co-located and launched together.

Also fixes a couple of build/runtime issues found while hardening this consolidated image, plus one pre-existing routing gap.

What Changed

Five Services, One Container (e748558)

scripts/start-combined.sh launches all five as separate uvicorn processes inside one container:

app            -> :8000
translation    -> :8001
field_mapping  -> :8002
ocr            -> :8010
gateway        -> :8080

If any one process dies, the script brings the whole container down (wait -n) rather than running silently degraded. This allows Docker Compose's restart policy to kick in instead of the container staying up in a half-broken state.

The root Dockerfile now installs all five services' dependencies. Each still has its own requirements.txt/pyproject.toml, installed as-is — this doesn't merge their code or dependency trees, it just installs all of them into one image.

docker-compose.yml drops from 5 backend service entries to 1 (backend/backend-gpu pair, same CPU/GPU opt-in pattern as ollama and surya-inference).

A new hf_cache volume persists the app's embedding model across container recreates instead of re-downloading it from HuggingFace every time.

Fixed a Pre-existing Routing Gap

The frontend called /cases through the gateway, but no route existed for it.

gateway/main.py now proxies /cases and /app/health to the app process, matching the existing OCR/translation/field-mapping proxy routes.

Ollama Ping Context-Size Mismatch (fca388f)

Ollama's num_ctx is fixed at llama-server process launch.

The translation health monitor (added in the previous branch) was pinging Ollama with different options than translate() uses (no num_ctx vs num_ctx=32768), which forced a full model reload (~250s) every time a ping and a translate call interleaved.

The monitor now reuses the same model_options for its ping, eliminating the reload thrashing.

Hardened the Consolidated Image (c2b9ac3)

• Excluded .env from the Docker build context — it was previously getting baked into image layers via COPY . ..
• Added restart: unless-stopped to db, ollama, and backend so a crash (which now brings down the whole combined container per the wait -n behavior above) actually restarts instead of staying down
• Removed the four now-dead per-service Dockerfiles (gateway, ocr, translation, field_mapping) — nothing builds from them anymore now that everything routes through the root Dockerfile
• Converted the root Dockerfile to a multi-stage build so gcc/libpq-dev and pip's build tooling stay in the build stage instead of shipping in the final runtime image

How to Test

cd lending-poc
docker compose up --build

1. Confirm docker compose ps shows 5 containers instead of the previous 9:

   • backend
   • db
   • ollama
   • surya-inference
   • frontend

2. Hit the gateway's proxied routes (/cases, /app/health, plus the existing OCR/translation/field-mapping paths) and confirm all five processes inside backend respond.

3. Kill one of the five uvicorn processes inside the backend container (e.g. docker exec + kill) and confirm the whole container exits and gets restarted by Compose, rather than continuing to run degraded.

4. Run a translate request followed by a health-check ping (or vice versa) back-to-back and confirm there's no ~250s stall from a model reload.

5. Run docker history (or inspect image size) on the built backend image to confirm gcc/libpq-dev aren't present in the final layer.

[Copilot]

🔵 Needs a closer look

It significantly changes container/build/runtime orchestration (multi-service co-location + Compose refactor), which warrants final human validation of startup, health, and restart behavior end-to-end.

Pull request overview

This PR consolidates five backend FastAPI services (app, translation, field-mapping, OCR, gateway) into a single Docker image/container while keeping them as independent uvicorn processes on their existing ports, and updates the local Compose stack accordingly. It also closes a gateway routing gap for /cases and aligns the translation service’s Ollama health ping options with real translate calls to avoid expensive model reload thrashing.

Changes:

• Add a combined entrypoint script to launch all five services in one container and exit the container if any process dies.
• Refactor docker-compose.yml + root Dockerfile for a single backend/backend-gpu service (multi-stage build, consolidated dependency installs, new hf_cache volume, restart policies).
• Update gateway routing to proxy /cases and /app/health, and fix translation Ollama ping to reuse translate-time model_options.

File summaries

File	Description
lending-poc/scripts/start-combined.sh	New entrypoint that starts five uvicorn processes and exits on first failure (wait -n).
lending-poc/README.md	Documentation updates for the consolidated backend container and new routing behavior.
lending-poc/gateway/main.py	Adds /cases and /app/health proxies; includes app in aggregated health.
lending-poc/gateway/Dockerfile	Removes now-dead per-service Dockerfile.
lending-poc/field_mapping_poc/Dockerfile	Removes now-dead per-service Dockerfile.
lending-poc/document_processing/translation/translation_service/adapters/ollama_adapter.py	Reuses model_options in ping to prevent Ollama model reload thrash.
lending-poc/document_processing/translation/Dockerfile	Removes now-dead per-service Dockerfile.
lending-poc/document_processing/ocr/Dockerfile	Removes now-dead per-service Dockerfile.
lending-poc/Dockerfile	Converts to multi-stage build and installs dependencies for all five services into a shared venv.
lending-poc/docker-compose.yml	Collapses multiple backend services into backend/backend-gpu, adds restart policies and hf_cache.
lending-poc/.env.example	Adds APP_REQUEST_TIMEOUT_SECONDS example.
lending-poc/.dockerignore	Updates ignores for the new consolidated build context and excludes .env.
.gitignore	Ignores CLAUDE.md.

Review details

• Files reviewed: 12/13 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.