← All projects

Hybrid RAG Knowledge Assistant

LLM chat assistant for an accounting practice that answers from Monday.com boards and OneDrive procedures, using BM25 + vector hybrid search with cited answers and a live retrieval trace.

Client
Accounting practice
Year
2024
Category
Full-Stack
Stack
  • Node.js
  • Claude API
  • transformers.js
  • BM25
  • Vector search
  • Server-Sent Events
  • Docker
Open live demo

The problem

An accounting practice kept everything in two places: Monday.com boards for clients, entities and about ten workflow boards (tax returns, BAS, payroll, TPAR…), and OneDrive for procedures and checklists. Staff spent time hunting for answers to questions like “what’s outstanding for this client?” or “what do we need before lodging this return?”.

What I built

  • Data pipeline: Monday.com webhooks and OneDrive sync fed MongoDB, classified tasks per client and per board, and re-embedded changed items automatically.
  • Hybrid search: keyword search for exact client codes and names, vector search for meaning (“paperwork” finds “awaiting client docs”), fused into one ranked list.
  • Grounded answers: the LLM answers only from the retrieved sources and cites each one. When the data doesn’t cover a question, it says so.
  • Follow-ups: “and who’s assigned?” is rewritten into a standalone search query before retrieval.

The live demo

The demo is a clean rebuild with invented data, so the client’s data stays private. It runs real retrieval: BM25, local MiniLM embeddings and weighted or RRF fusion. Claude Haiku streams the answers. Each answer shows its retrieval trace: query analysis, keyword hits, vector hits, fusion ranks and the exact context sent to the model. A slider moves the search from pure keyword to pure semantic.

The public endpoint is protected against abuse: signed browser cookies, an origin check, limits per browser and per IP, and global daily caps on messages and tokens.