Munich Intel

An AI research assistant for Munich startups — scrapes company pages, embeds them with BGE-M3, and answers questions using RAG.

Live demo: https://huggingface.co/spaces/anushkasinghh/munich-intel

Stack

• Groq — LLM inference (llama-3.1-8b-instant, cloud API)
• Qdrant Cloud — hosted vector database (production) / local Docker (dev)
• BGE-M3 via sentence-transformers — dense embeddings
• FastAPI — API + streaming chat UI
• uv — dependency management and lockfile
• HuggingFace Spaces — deployment (Docker, free CPU tier)

Project Structure

munich-intel/
├── .github/
│   └── workflows/ci.yml        ← lint (ruff) + unit tests on every push
├── data/
│   └── raw/                    ← scraped JSON files, gitignored
├── src/
│   └── munich_intel/
│       ├── config.py           ← all settings, loaded from .env
│       ├── scraper.py          ← fetches and cleans HTML
│       ├── chunker.py          ← splits text into indexable pieces
│       ├── embedder.py         ← wraps BGE-M3
│       ├── indexer.py          ← writes to Qdrant
│       ├── retriever.py        ← queries Qdrant
│       ├── generator.py        ← calls Groq API
│       └── pipeline.py         ← retriever + generator = RAG answer
├── api/
│   └── main.py                 ← FastAPI: /ingest, /query, /query/stream, /health
├── scripts/
│   └── ingest.py               ← CLI: python scripts/ingest.py --company twaice
├── tests/
│   ├── test_chunker.py         ← invariant unit tests (CI)
│   ├── test_embedder.py        ← integration tests, loads real BGE-M3 (manual only)
│   └── test_pipeline.py        ← wiring contract tests with mocks (CI)
├── companies.yaml              ← data source list, not code
├── DECISIONS.md                ← architecture decision log
├── docker-compose.yml          ← Qdrant only
├── pyproject.toml
└── .gitignore

Design Decisions

Data sources in YAML, not code — adding a company means editing companies.yaml, not Python.

src/ layout — prevents Python from accidentally importing local files instead of installed packages. Standard for anything you might eventually package or deploy.

api/ separate from src/ — the FastAPI layer is a delivery mechanism, not business logic. Keeping it outside src/ makes that boundary explicit.

Config via environment variables — src/munich_intel/config.py uses Pydantic BaseSettings. Change any value via .env or environment variable, never by editing code.

Phase 1: dense-only search — sentence-transformers wraps BGE-M3 for dense vectors. Hybrid search (dense + sparse) comes in Phase 2 with FlagEmbedding.

See DECISIONS.md for the full rationale on each choice.

Setup

Prerequisites

• Docker (for local Qdrant)
• A Groq API key — free at console.groq.com
• Note: BAAI/bge-m3 (~570MB) downloads from HuggingFace on first run

Install

uv sync --extra dev

Configure

cp .env.example .env
# Fill in GROQ_API_KEY at minimum. Defaults work for local Qdrant dev.

Run Qdrant (local dev)

docker compose up -d

Start the API

uv run uvicorn api.main:app --reload

Ingest companies

The /ingest endpoint requires the X-Ingest-Token header. Generate a secret once and set it as INGEST_SECRET in .env and in HF Space secrets:

# Generate a secret
python3 -c "import secrets; print(secrets.token_hex(32))"

# Trigger ingest (all companies)
curl -X POST http://localhost:8000/ingest \
  -H "X-Ingest-Token: your-secret" \
  -H "Content-Type: application/json" \
  -d '{}'

# Or a single company
curl -X POST http://localhost:8000/ingest \
  -H "X-Ingest-Token: your-secret" \
  -H "Content-Type: application/json" \
  -d '{"company_slug": "reverion"}'

Running Tests

# Unit tests (fast, run in CI)
uv run pytest tests/test_chunker.py tests/test_pipeline.py -v

# Integration tests (loads real BGE-M3 model, ~15s — run manually before model swaps)
uv run pytest tests/test_embedder.py -v -m integration

Phase 1 (MVP)

Dense vector search over scraped Munich startup content. Endpoints: POST /query, POST /query/stream, POST /ingest (auth required), GET /health.

Phase 2 (Post-MVP)

Hybrid search (dense + sparse vectors) using FlagEmbedding. See DECISIONS.md.