fungzyui/hkex-director-rag

HKEX Director RAG

Hybrid retrieval-augmented generation over 73,305 HKEX director-appointment records (~3,569 listed companies, ~43,529 directors).

Three retrievers fused under an LLM router:

1. 1.SQL (DuckDB) — exact lookups, aggregates, date filters.
2. 2.BM25 — keyword recall with an EN+CJK tokenizer (handles 范明 / 李嘉誠).
3. 3.Dense — intfloat/multilingual-e5-small embeddings in FAISS.

The router uses OpenAI-compatible function calling against OpenRouter (default openai/gpt-oss-120b:free) and picks which retriever(s) to call.

Usage

• —🔎 Browse and 📊 Stats tabs work as-is.
• —💬 Ask tab requires an OpenRouter key — paste it into the sidebar (per-session, not stored). Get one at openrouter.ai/keys.

Source

Full code, architecture diagram, and design notes: github.com/fungzyui/hkex-director-rag.