ASB Secure Gateway

Minimal FastAPI + OPA reference that demonstrates:

• OpenAI-compatible /v1/chat/completions proxy guarded by prompt policies.
• /v1/rag/search_safe endpoint backed by Postgres + pgvector with safety checks.
• /v1/agent/action/execute minimal agent tool executor with policy enforcement.
• ASB Security Schema events shipped to OPA as input for every operation.
• Single docker-compose stack (API + Postgres/pgvector + OPA).

Stack

• Python 3.11, FastAPI, Uvicorn
• Postgres pgvector/pgvector:pg16 demo corpus (docker/init/01_init.sql)
• Open Policy Agent sidecar loading policies/*.rego

Running the stack

1. Export environment variables (or create an .env) - defaults already match docker-compose:

   # PowerShell examples; replace with export for other shells
   setx OPA_URL http://opa:8181
   setx DATABASE_URL postgresql://postgres:postgres@postgres:5432/asb_gateway
   setx OPENAI_API_KEY ""   # optional, leave empty for demo responses

2. Build + start everything:

   docker-compose up --build

3. API is available at http://localhost:8000 (Swagger UI at /docs), OPA at http://localhost:8181, Postgres at localhost:5432.

Key endpoints

Endpoint	Description	Notes
POST /v1/chat/completions	OpenAI-compatible chat completions	Streams disabled; forwards to OpenAI when OPENAI_API_KEY is set, otherwise returns deterministic demo text.
POST /v1/rag/search_safe	Secure RAG search	Queries pgvector table documents with dimensionality 6. Falls back to demo data if DB unavailable.
POST /v1/agent/action/execute	Agent action gateway	Restricts execution to AGENT_ALLOWED_TOOLS env var (defaults: ping,whoami).
GET /health	Health probe	Returns { "status": "ok" }.

All routes emit ASB Security Schema v0.1 events (subject / operation / resource / context / decision) and expect an allow decision from OPA under:

• data.prompt.allow
• data.rag.allow
• data.agent.allow

See policies/*.rego for the minimal guardrails (temperature caps, top_k limits, denied tools, etc.). Modify those files and OPA will live-reload thanks to the shared volume.

Local development

pip install -r requirements.txt
uvicorn app.main:app --reload

Set DATABASE_URL to point at a Postgres instance with pgvector + the documents table. The dockerized Postgres already includes sample rows and the required vector extension.

Project layout

app/
  config.py        # Pydantic settings
  main.py          # FastAPI entrypoint + lifespan hooks
  opa_client.py    # Async HTTP client used by services
  models/          # ASB event + request/response schemas
  routes/          # Thin FastAPI routers per feature
  services/        # Policy-aware service layer
policies/          # prompt/rag/agent rego policies
docker/init/       # pgvector bootstrap SQL

Useful environment variables

Variable	Default	Description
OPA_URL	http://opa:8181	Location of the OPA server
DATABASE_URL	postgresql://postgres:postgres@postgres:5432/asb_gateway	Postgres DSN for RAG search
OPENAI_API_KEY	""	Optional upstream OpenAI API key
OPENAI_BASE_URL	https://api.openai.com	Overridable base URL
AGENT_ALLOWED_TOOLS	ping,whoami	Comma-separated list of allowed agent tools

Testing the APIs

# Chat (demo response unless OPENAI_API_KEY is supplied)
curl -s http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"Hello"}]}'

# RAG search
curl -s http://localhost:8000/v1/rag/search_safe \
  -H "Content-Type: application/json" \
  -d '{"query":"What is the gateway?"}'

# Agent tool
curl -s http://localhost:8000/v1/agent/action/execute \
  -H "Content-Type: application/json" \
  -d '{"tool":"ping"}'

Looking for commercial support, enterprise-specific rules, or expanded policies? Reach out via GitHub discussions or email the maintainer to discuss commercial licensing and premium rule packs.