LLMs4EU - Tourism - RAG

Local RAG for tourism places. SQLite stores canonical content, and Chroma stores derived vector indexes. Both run entirely inside the Python environment.

Requirements

• uv
• just
• Ollama — only needed for just ask, just scrape, and just eval-generate

Setup

uv sync --extra dev   # install all dependencies
cp .env.example .env  # set local paths (defaults work out of the box)

ollama pull gemma4:e4b
# Optional for eval question generation:
ollama pull gemma4:26b-a4b-it-q4_K_M
ollama pull gpt-oss:20b

Usage

just init        # load seed data into SQLite
just index       # embed places and rebuild Chroma
just ask What place is best for a quiet forest walk near water?  # retrieve + LLM answer
just scrape https://example.com  # crawl a site, ingest into SQLite, reindex
just scrape-web  # start the local scraping UI
just eval-index qwen  # rebuild the default qwen chunk vector index
just test        # run the non-LLM test suite

Search defaults to top 10 final results. Override per query:

uv run python -m src.rag.search "lake picnic" --limit 5

Experimental retrieval eval over scraped Markdown pages:

just eval-chunks
just eval-index qwen
just eval-generate 10
just eval qwen,sparse
just eval qwen4b_rerank,qwen4b_hybrid,qwen4b_hybrid_rerank

Shape

data/           tracked seed fixture
sql/            one-table schema, portable to SQLite and Postgres
src/db/         SQLite initialize and place queries
src/preprocess/ rebuild derived data from SQL rows
src/indexing/   provider-shaped vector indexing
src/vector_store/  Chroma collection, upsert, vector search
src/retrieval/  chunk retrieval methods and catalog
src/rag/        place search and answer scripts
src/eval/       chunked raw-page retrieval evaluation
src/scraping/   crawler, transform, ingest, scraping UI
src/shared/     schema, embeddings, env, LLM helper
tests/          data contract, retrieval, scrape transform

Durable reference databases live under data/db/ with descriptive names such as pages.db. Regenerable vector cache artifacts live under data/cache/chroma/. SQLite page chunks are the source of truth for chunk text; Chroma collections are derived indexes over those chunks. Use .env overrides for private scratch paths under .local/.

---

Roadmap

This repo focuses on building a clean, portable place database as the foundation for a larger RAG system.

Current: SQLite + Chroma, everything local, no services needed.

Next step: swap storage backends for larger shared runs:

• SQLite → Supabase (free tier, under 500 MB, shared across teams)
• Chroma → Qdrant (free tier, managed vector index)

The SQL schema and module boundaries are designed to make that swap small. Before scaling up, we need to align with other teams on what data collection tools and shared infrastructure are available.

Docs

• docs/architecture-decisions.md: durable decisions and why they matter.
• experiments/indexing/README.md: retrieval experiments and evaluation protocol.
• docs/retrieval-results-agentic.md: current agentic retrieval report.