RAG-Scheduler

The problem

Timetables in spreadsheets are hard to query quickly. Asking “What do I have on Tuesday afternoon?” should just work — without sending personal data to a cloud API.

What it does

• Upload a CSV; each row is embedded once with all-MiniLM-L6-v2 and indexed in HNSWLib.
• Queries retrieve the top-k matching rows and pass them to a local LLM via Ollama as grounded context.
• REST API with /status, /ingest and /ask; answers return the source rows they were based on.
• Runs fully offline — no API keys, no per-call cost, no data leaves the machine.

Architecture

• Frontend: React + Vite with a custom useChat hook owning chat state and a services/api.js layer for backend calls.
• Backend: Express routes stay thin; services handle embeddings, vector storage, CSV ingestion and the RAG chain.
• The vector index is persisted to disk so ingestion only happens once per timetable.

Method

• Embed the user's question with the same model used at ingest time.
• Return the most similar rows by cosine distance from HNSWLib.
• Build a prompt that constrains the model to answer only from the retrieved rows.

Lessons and next steps

• The backend was rewritten from FastAPI to Node.js without touching the frontend, because the REST contract was defined first.
• Next steps documented in the repo: streaming responses, chat history persistence and PDF/calendar ingestion.

Built with

• React
• Vite
• Node.js
• Express
• Transformers.js
• HNSWLib
• Ollama
• Multer