feat(chat): multi-model /tasks/chat/completions (sunflower-14b + sunflower-9b)

[PatrickCmd]

Summary

Extends the OpenAI-compatible POST /tasks/chat/completions endpoint from a single hardcoded model to two Sunflower text models, each on its own RunPod serverless endpoint. Backward compatible with existing string-content clients; also documents the endpoint across the API docs, tutorial, and frontend, and wires the new config into the Cloud Run deploy.

model	Endpoint env var	Coverage
sunflower-14b (default)	SUNFLOWER_14B_ENDPOINT_ID (falls back to SUNFLOWER_ENDPOINT_ID / QWEN_ENDPOINT_ID)	English + 31 Ugandan/regional languages (32 total)
sunflower-9b	SUNFLOWER_9B_ENDPOINT_ID	67 African languages (good/moderate/basic tiers)

Behavior

• Model selection via the model field; sunflower-14b is the default when omitted. Unknown model → 400 listing supported models.
• No alias. The old Sunbird/Sunflower-14B identifier is not accepted — it returns 400 telling the client to use sunflower-14b.
• Text only. Message content is a plain string; non-string content → 422.
• Upstream served name. The model= forwarded to each RunPod/vLLM endpoint defaults to the short name (sunflower-14b / sunflower-9b) to match each endpoint's --served-model-name; overridable via SUNFLOWER_14B_MODEL_NAME / SUNFLOWER_9B_MODEL_NAME.
• Quota/billing unchanged. Legacy internal qwen/sunflower keys are preserved and now also send the short upstream name (the translation service routes through them to the same 14B endpoint).

Scope note: An earlier revision of this branch added Sunflower Gemma 2B with base64 audio input. That was removed before deployment by product decision — the endpoint serves only the two text models above. Branch history preserves it; the design spec carries an amendment note.

Changes

Backend

• app/schemas/chat.py — SUPPORTED_MODELS = ("sunflower-14b", "sunflower-9b"), DEFAULT_MODEL, RENAMED_MODELS (drives the deprecated-identifier error), resolve_model (no aliases), string-only content.
• app/services/inference_service.py — env-driven multi-endpoint routing map for both models plus legacy qwen/sunflower keys; short upstream served names.
• app/routers/chat.py — per-model routing (no hardcoded model constant); _validate_model returns the targeted rename error or the generic unsupported-model error.

Docs

• app/docs.py, docs/tutorial.md, frontend/src/pages/Tutorial.tsx — model selection, per-model language coverage (sourced from the SALT model pages / HuggingFace), removed-alias behavior, and updated code samples. Rebuilt SPA assets committed.

Deployment / config

• .github/workflows/deploy-api.yml — adds SUNFLOWER_14B_ENDPOINT_ID / SUNFLOWER_9B_ENDPOINT_ID to the Cloud Run deploy via env_vars_update_strategy: merge (preserves the service's existing env). Azure/Heroku untouched.
• .env.example — documents the new endpoint vars, the 14B fallback chain, and the optional model-name overrides.

Testing

• Full suite green apart from 4 pre-existing Google-Analytics-config failures (present at the base commit, unrelated). New coverage: per-model routing, default model, deprecated-alias rejection (400), unknown model (400), non-string content (422), upstream served-name defaults, and legacy-key regression; existing string-content/streaming tests still pass.
• app/ is black/isort/flake8-clean; frontend npm run build (tsc + vite) passes.

Ops before / after deploy

• Set the GitHub Actions secrets SUNFLOWER_14B_ENDPOINT_ID and SUNFLOWER_9B_ENDPOINT_ID before the next Cloud Run deploy. (14B still falls back to the existing SUNFLOWER_ENDPOINT_ID/QWEN_ENDPOINT_ID on the service; 9B needs its secret to function.)
• If any endpoint is redeployed under a different --served-model-name, set SUNFLOWER_14B_MODEL_NAME / SUNFLOWER_9B_MODEL_NAME on the service to match.

Known, out of scope

• npm run lint fails due to a pre-existing ESLint v10 / flat-config toolchain mismatch (fails identically without this branch's changes); the npm run build type-check gate passes.

🤖 Generated with Claude Code