EdgeSLM Smart Store Assistant

Voice-enabled, offline laptop recommender that runs a swappable Small Language Model locally on a Raspberry Pi 5 (or any Linux box). FAISS-backed RAG over a 25-laptop INR catalogue, rule-based clarification, optional Whisper/Piper voice I/O, browser UI on LAN.

See edge-slm-smart-store.md (full design doc) and PROGRESS.md (implementation status).

Quick start (development)

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt            # or: pyyaml numpy fastapi uvicorn pydantic sentence-transformers faiss-cpu python-multipart
python scripts/build_index.py              # builds FAISS index from catalogue
uvicorn edgeslm.main:app --app-dir src --host 0.0.0.0 --port 8080

Open http://localhost:8080/ (user) or http://localhost:8080/admin (admin).

If no GGUF is present the backend falls back to a deterministic template response — the app is fully usable for retrieval + dialogue without any downloaded models. To enable the SLM and voice:

bash scripts/download_models.sh            # Qwen2.5-1.5B + Whisper tiny.en + Piper

Tests

pytest -q                                  # 37 tests, <1s

Deploy to Raspberry Pi 5

bash scripts/install_pi.sh                 # on the Pi
sudo cp systemd/edgeslm.service /etc/systemd/system/
sudo systemctl enable --now edgeslm

Layout

configs/          app.yaml + models.yaml (single source of truth)
data/catalogue/   laptops.json (25 items)
data/index/       persisted FAISS index + embeddings
src/edgeslm/
  catalogue/      Laptop dataclass + JSON/CSV loader
  retrieval/      MiniLM + FAISS, pre-filter, pool-restricted search
  dialogue/       regex parser, rule clarifier, session store
  generation/     llama.cpp wrapper, grounded RAG prompt, guard, fallback
  voice/          whisper.cpp STT, Piper TTS, Bhashini stub
  api/            FastAPI routes (chat / voice / admin) + pydantic schemas
  orchestrator.py single boundary: handle_turn(session_id, text)
  main.py         FastAPI app + WebSocket token streaming + SPA hosting
scripts/          build_index, download_models, bench, install_pi, start
tests/            pytest suite (catalogue, retrieval, parser, clarifier, prompt, e2e)
web/dist/         minimal SPA (index.html = user, admin.html = admin)