Nutrition Research Assistant

Production-grade multi-agent Indian nutrition research assistant: answers questions about Indian food and nutrition by combining a curated internal knowledge base, hybrid RAG, specialized agents, MCP tools, conversational memory, optional live web search, graceful failure handling, and end-to-end OpenTelemetry-compatible tracing.

graph TD
    Client[CLI / Swagger / Demo script] --> API[FastAPI :8000]
    API --> Sup[Agno Supervisor]
    Sup --> KA[Knowledge Agent]
    Sup --> CA[Calculator Agent]
    Sup --> DA[Document Agent]
    Sup --> WA[Health/Web Agent]
    KA --> MCP[FastMCP Server :8001/mcp]
    CA --> MCP
    DA --> MCP
    WA --> MCP
    MCP --> Chroma[ChromaDB + BM25 + RRF + Reranker]
    MCP --> Calc[Safe calculator]
    MCP --> Web[Agno DDGSTools]
    MCP --> Docs[documents.json]
    Sup --> Mem[(SQLite memory)]
    API --> OTel[OpenTelemetry + structlog]

Core principle (PRD §89): the LLM decides what should happen; MCP tools perform what must happen deterministically. Agents never touch Chroma directly — everything flows through MCP.

Tech Stack

Related MCP server: Food Data Central MCP Server

Repository Layout

corpus/         8 curated nutrition documents (frontmatter + per-food tables)
scripts/        ingest.py · rebuild_index.py · test_retrieval.py · eval_questions.json
                test_mcp_wiring.py · test_agents.py · test_memory.py · demo.py
src/
  api/          FastAPI app, routes, schemas          (chat / health / ready)
  agents/       models.py (DeepSeek) · mcp_client.py · agents.py (specialists + supervisor)
  mcp_server/   server.py + tools/{knowledge,calculator,documents,web}.py
  rag/          embeddings · chroma · bm25 · fusion (RRF) · reranker · pipeline
  memory/       SQLite storage
  services/     circuit breaker · retry
  observability tracing (OTel) · logging (structlog)
  config/       settings.py          core/ errors.py · models.py
  cli.py        REPL client
tests/          unit + API tests

Quick Start

# 1. Environment
uv venv --python 3.12
uv sync

# 2. Secrets
cp .env.example .env        # set DEEPSEEK_API_KEY

# 3. Ingest the corpus (downloads bge-small-en-v1.5 on first run)
uv run python scripts/ingest.py

# 4. Start the MCP tool server  (terminal 1)
uv run python -m src.mcp_server.server

# 5. Start the API                 (terminal 2)
uv run uvicorn src.api.app:app --reload --port 8000

Swagger UI: http://127.0.0.1:8000/docs (API port is configurable via API_PORT in .env — this workspace uses 8002 because 8000 is taken)

Try it:

uv run python -m src.cli                # interactive REPL
uv run python scripts/demo.py           # scripted 5-question demo
curl -X POST http://127.0.0.1:8000/api/v1/chat \
  -H "Content-Type: application/json" \
  -d '{"session_id":"demo","message":"How much protein is in 100g cooked chickpeas?"}'

API

Degraded responses carry status: "degraded", a human-readable message, and the trace_id — never a 500 traceback.

MCP Tools (port 8001)

Agents

Failure Handling

• MCP down → controlled degraded response with trace_id (no 500); circuit breaker fails fast, probes after cooldown (PRD §70/§85)

• MCP timeout — 10s transport read timeout; agent run bounded by AGENT_TIMEOUT_SECONDS

• Chroma down → "Knowledge retrieval temporarily unavailable" (optionally falls back to web)

• LLM down / bad key → LLM_UNAVAILABLE-style degraded message

• Web disabled → "Live search is currently unavailable."

• Retry policy: 2 attempts with backoff, then degrade — never endless

Observability

Every request produces one trace: api.request → supervisor.run → mcp.<tool> → rag.dense_search / rag.bm25 / rag.rrf / rag.reranker, with trace_id echoed in the API response, logs, and spans. All logs are JSON (structlog) with timestamp, level, logger, trace_id, span_id. Set OTEL_EXPORTER_OTLP_ENDPOINT to export to Jaeger/Tempo/Collector (uv sync --extra otel).

Tests & Retrieval Eval

uv run pytest -q
uv run python scripts/test_retrieval.py --rerank   # Recall@5/10, MRR, Hit@1/3 over 24 questions

The eval dataset (scripts/eval_questions.json, 24 questions across 7 categories) measures retrieval quality; the reranker lifts Hit@1/3 over raw fusion. Results land in data/eval/results.json.

Definition of Done (PRD §87)

• Core — FastAPI runs; DeepSeek via OpenAI-compatible client; Agno agents operational; FastMCP operational; Chroma populated; 8 documents; memory works

• MCP — all 5 tools work and agents use MCP as the real execution path

• RAG — dense + BM25 + RRF + cross-encoder rerank; metadata preserved; eval exists

• Agents — Supervisor + 4 specialists; multi-agent composition works

• Memory — session ids; follow-ups resolve; persisted in SQLite

• Observability — trace_id per request; nested spans; MCP/RAG/LLM/errors traced

• Failure handling — MCP unavailable/timeout, Chroma, LLM, web all controlled; no unhandled exception reaches the user

Troubleshooting

• Hugging Face symlink warning — Windows-only; set HF_HUB_DISABLE_SYMLINKS_WARNING=1 (cosmetic)

• Retrieval returns nothing — Chroma empty: run python scripts/ingest.py

• "tool service unavailable" — MCP server not running (start it) or breaker cooling down (wait ~30s, or restart API)

• Health says llm not_configured — DEEPSEEK_API_KEY missing in .env