Brief #12 · AI & LLM
RAG Chatbot for Your Docs
Medium⏱ 12-16 hSoon
A chatbot that answers questions from your company's documents, cites its sources, and says 'I don't know' when it should.
Outlines are free · building with tracking needs Pro
RThe briefFrom the team
New hires ask HR the same thirty questions about leave, payroll and policies, all answered somewhere in a 200-page handbook nobody reads. Build an assistant that answers from the handbook only, with a link to the exact section, and never makes up a policy.
What you'll build
- 1Ingest PDFs and Markdown, split them into chunks with overlap, and keep each chunk's source and heading.
- 2Embed the chunks and store them in a vector index (pgvector, Qdrant or similar).
- 3Retrieve the top chunks for a question and answer with citations.
- 4Refuse when retrieval finds nothing relevant, instead of guessing.
- 5A test set of 30 questions with expected sources, to measure retrieval quality.
Step-by-step walkthrough
Walkthrough coming soon
This project's step-by-step walkthrough, with real outputs at every step, is still being written. You can read the brief and try it on your own meanwhile; tracking opens when the walkthrough is ready.
Done when
- Every answer links to the chunks it used.
- On your test set, the right source is retrieved for at least 85% of questions.
- Questions outside the handbook get an honest 'not in the documents'.
- A README compares two chunk sizes with numbers.
Stretch goals
- Hybrid search: combine keyword (BM25) and vector results.
- Bangla questions over English documents.