Build
A document assistant that cites its sources
Retrieval-augmented generation over internal documents, built around one rule: every answer must point to the paragraph it came from.
- Role
- Product & engineering
- Year
- 2024 — Present
- Industry
- Enterprise
- Status
- In progress
01The challenge
People spent hours searching policies and contracts. Generic chatbots answered confidently and wrongly.
02The approach
Chunking that respects document structure, hybrid search (full-text + vectors) in PostgreSQL, answers constrained to retrieved passages, and an evaluation set that runs on every change.
03Architecture
Ingestion pipeline → PostgreSQL with pgvector → hybrid retriever → LLM with citation-enforcing prompt → evaluation harness.
Answering with evidence
Ingest
Parse
Chunk by structure
Embed
Retrieve
Full-text
Vector
Re-rank
Answer
LLM
Citations
Evaluation
Outcome
Status and early results