The AI planning microservice for the Matrix EcoSystem. It generates short, low-risk, auditable remediation plans for Matrix-Guardian and offers a compact, RAG-assisted chat over Matrix docs.
Given a compact health context from Matrix-Guardian, matrix-ai produces a JSON plan with bounded steps, a risk label, and a brief rationale. Strict schema, PII redaction, short timeouts, and retries ensure safe operation.
Lightweight retrieval over curated Matrix docs. We assemble a compact CONTEXT and instruct the model to answer only from those facts (or say it doesn’t know).
Production-safe defaults: JSON logs, request IDs, in-memory rate-limit, idempotent writes (where applicable), and ETag support for caches.
POST /v1/plan # JSON plan for Guardian (non-stream)
POST /v1/chat # Q&A over Matrix docs (RAG, stream or non-stream)
GET /healthz # Liveness# 1) Export your token (or use Space Secret)
export HF_TOKEN="hf_xxx"
# 2) Run the service
uvicorn app.main:app --host 0.0.0.0 --port 7860
# 3) Try chat (non-stream)
curl -s -X POST localhost:7860/v1/chat \
-H 'content-type: application/json' \
-d '{"query":"What is MatrixHub?"}' | jq
# 4) Try plan
curl -s -X POST localhost:7860/v1/plan \
-H 'content-type: application/json' \
-d '{"mode":"plan","context":{"entity_uid":"matrix-ai"},"constraints":{"max_steps":3,"risk":"low"}}' | jq