clossers/didi-lot1-ai
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
didi-lot1-ai/ai_platform/modules/catalog-api/INDEX.md

8.8 KiB
Raw Blame History

catalog-api — INDEX

Catalog API for the DIDI AI platform. It is a service registry and discovery gateway: it aggregates /v1/info responses from each ML component (LLM, Audio, Video, Web) and re-exposes a unified view (components, models, functions, status) plus a merged OpenAPI 3.1 spec with Swagger UI / ReDoc. Production port 11000, container didiAI-catalog-api. Read-only HTTP aggregator — no database, no writes.

  • Stack: Python 3.11, FastAPI, httpx (async), pydantic-settings, uv, uvicorn
  • Container: didiAI-catalog-api
  • Internal URL: http://didiAI-catalog-api:11000
  • Production host: http://10.11.10.42 (per CATALOG_EXTERNAL_URL in deploy/.env.example)
  • Network: Docker external network didi-network (shared with the other ai_platform modules)
  • Sister CLAUDE.md (platform): /home/admin365/didi_mono/ai_platform/CLAUDE.md

Note: despite the name, this module is not a knowledge-graph / atom catalog. It is a service catalog (think "service registry" in the microservices sense). It does not talk to PostgreSQL, Atomic, Redis, or didi-brain.

Ce face

  • Aggregates static metadata from each ML component by calling GET /v1/info on the configured backends.
  • Returns a single unified response with:
    • Resources — component descriptor (name, slug, resource_type, ...).
    • Models — every model exposed by every backend (LLM, Whisper, vision, etc.).
    • Functions — every endpoint/function each backend advertises.
  • Probes liveness of every backend and reports aggregated health (healthy / degraded / unhealthy).
  • Builds a merged OpenAPI 3.1.0 document (paths prefixed with /{component_id}, schemas prefixed with {component_id}_) and serves it as JSON, Swagger UI, and ReDoc — so a frontend or backend can consume a single contract for all GPU services.
  • Tolerant to partial outages: if a component is unreachable it is recorded under errors and skipped, the rest of the catalog still serves.

API endpoints

Source: src/catalog_api/app.py and API.md.

Method Path Description
GET /health Liveness probe ({"status":"ok"}).
GET /v1/components List all registered components with full metadata (resource + models + functions).
GET /v1/components/{component_id} Full metadata for one component (404 if unknown).
GET /v1/models Flattened list of every model across components.
GET /v1/functions Flattened list of every function/endpoint across components.
GET /v1/status Aggregated reachability + health of all components (healthy_count/total_count).
GET /v1/openapi Merged OpenAPI 3.1 spec (all components, prefixed).
GET /v1/docs Swagger UI for the merged spec.
GET /v1/redoc ReDoc for the merged spec.
GET /v1/openapi/component/{component_id} Raw OpenAPI spec of a single component (passthrough).

Status values from /v1/status: healthy (all reachable), degraded (some reachable), unhealthy (none reachable).

When accessed through the gateway, Bearer-token auth is enforced by nginx (<host>:11000/catalog/); direct access is unauthenticated by design.

Architecture

+-----------------------------------------------------+
|  Catalog API (:11000)  — didiAI-catalog-api         |
|  - Calls /v1/info on each backend                   |
|  - Merges OpenAPI specs, exposes Swagger UI / ReDoc |
|  - Stateless, no DB                                 |
+-------+----------+----------+----------+------------+
        |          |          |          |
        v          v          v          v
   LLM 14011  Audio 54300  Video 54600  Web 51100
   didiAI-llm-api  -audio-api  -video-api  -web-api
  • Pure aggregator — no persistence, no caching layer; a fresh fan-out happens per request via httpx.AsyncClient.
  • Component list is config-driven (CatalogSettings.get_components() in settings.py): a component with an empty URL is silently dropped (used today to disable Video by setting CATALOG_VIDEO_URL="").
  • External-vs-internal URL split — internal URLs (*_URL) are used for live calls inside the Docker network; the CATALOG_EXTERNAL_URL + *_EXTERNAL_PORT pair is the public base URL injected into the merged OpenAPI servers: so that external clients hit the right hostnames/ports.

How it is consumed

  • Frontend / API gateway — fetches /v1/openapi to expose Swagger UI for the whole platform; fetches /v1/status for a system-health widget.
  • Backend integrations — pull /v1/components to populate their own catalog tables (catalog.resources, catalog.models, catalog.functions), as illustrated in README.md § Use Cases.
  • Service discovery — clients query /v1/models to find a model by model_type (e.g., all vision models) without hard-coding hosts.

Structura fișiere

catalog-api/
├── README.md              Overview, quick-start, configuration, use cases
├── API.md                 Endpoint reference (request/response shapes)
├── INDEX.md               This file
├── pyproject.toml         Hatchling package, FastAPI/httpx/pydantic-settings deps
└── src/catalog_api/
    ├── __init__.py        version = "0.1.0"
    ├── app.py             FastAPI app — all endpoints + OpenAPI merger (~21 KB, single module)
    └── settings.py        CatalogSettings (env prefix CATALOG_) and Component model
└── deploy/
    ├── Dockerfile         Multi-stage build: python:3.11.12-slim + uv 0.10, runs uvicorn
    ├── docker-compose.yml Defines didiAI-catalog-api on didi-network, expose:11000 only
    ├── deploy.sh          Wrapper: loads .env, validates CATALOG_EXTERNAL_URL, runs compose
    ├── .env.example       Documented environment variables
    └── .env               Local environment (CATALOG_EXTERNAL_URL=...)

Implementation footprint is tiny: one app.py (all endpoints + OpenAPI merger live there) plus one settings.py.

Configuration

All settings come from environment variables with prefix CATALOG_ (see src/catalog_api/settings.py).

Variable Default Purpose
CATALOG_HOST 0.0.0.0 Bind address.
CATALOG_PORT 11000 Bind port.
CATALOG_LOG_LEVEL INFO Python logging level.
CATALOG_EXTERNAL_URL required Public base URL (e.g., http://10.11.10.42) injected into the merged OpenAPI servers:. deploy.sh aborts if missing.
CATALOG_LLM_URL http://didiAI-llm-api:14011 LLM Inference internal URL.
CATALOG_AUDIO_URL http://didiAI-audio:54300 Audio API internal URL.
CATALOG_VIDEO_URL (empty) Video Analysis internal URL — empty string disables Video.
CATALOG_WEB_URL http://didiAI-web-api:51100 Web API internal URL.
CATALOG_LLM_EXTERNAL_PORT 14011 External port advertised in the merged OpenAPI for LLM.
CATALOG_AUDIO_EXTERNAL_PORT 54300 External port for Audio.
CATALOG_VIDEO_EXTERNAL_PORT 54600 External port for Video.
CATALOG_WEB_EXTERNAL_PORT 51100 External port for Web.
CATALOG_COMPONENT_TIMEOUT 10 Per-call httpx timeout in seconds.

Deployment

  • Docker compose (deploy/docker-compose.yml): builds didiai-catalog-api, attaches to external network didi-network, only expose: 11000 (no host port — traffic comes through the platform gateway). Healthcheck hits http://localhost:11000/health every 30 s, restart policy unless-stopped.
  • Dockerfile (deploy/Dockerfile): two-stage build using ghcr.io/astral-sh/uv:0.10 for dependency install, then a slim python:3.11.12-slim runtime that runs python -m uvicorn catalog_api.app:app.
  • Helper script (deploy/deploy.sh): loads .env, validates CATALOG_EXTERNAL_URL, supports --detach, --down, --logs.
  • Local dev (per README.md): uv sync && uv run python -m uvicorn catalog_api.app:app --host 0.0.0.0 --port 11000 (requires the listed components reachable on the network).

Related modules

This service stands on top of the rest of the ai_platform/modules/* family — they are its data sources:

  • llm-inference (port 14011, container didiAI-llm-api) — chat, embeddings, rerank.
  • audio (port 54300, container didiAI-audio) — transcription / TTS.
  • video-analysis (port 54600, container didiAI-video-api) — vision pipelines (currently disabled by default in the .env example).
  • web (port 51100, container didiAI-web-api) — fact-check / web crawler.
  • dashboard — primary frontend consumer of the merged OpenAPI / /v1/status.

It is independent of:

  • didi-brain, Atomic / knowledge-graph services, PostgreSQL, Redis — none of these are accessed.
  • The orchestration-layer (agent-v3) does not currently consume this catalog; it talks to the GPU services directly.