Project
Production systemDocument Intelligence (RAG)
AI retrieval · Internal knowledge
Retrieval-augmented search over large internal documents with ranking, citations, and LLM-grounded answers.
Problem
Internal knowledge lived in long PDFs; keyword search produced inconsistent, untraceable answers.
Solution
RAG pipeline with hybrid retrieval, LLM answers with citations, and a web UI for querying sources.
Result
Lookup time dropped from minutes to seconds, with a single trustworthy interface for decisions.
What it does
- Ingestion pipeline: chunking, metadata, and embeddings
- Hybrid retrieval tuned for domain terminology
- LLM answers with citations and guardrailed prompts
- Web UI for queries, follow-ups, and inspecting sources
Overview
Internal knowledge lived in long PDFs and policy docs; keyword search and ad-hoc ChatGPT paste-ins produced inconsistent, untraceable answers. The product direction was a trustworthy assistant: upload or index documents, ask questions in natural language, and get answers tied back to sources.
I built the ingestion side—chunking, metadata extraction, and embeddings—then retrieval with hybrid search to handle domain vocabulary, an LLM layer that formats answers with citations, and guardrails around prompts and outputs. The web UI supports follow-up questions and exploring retrieved chunks so users can verify what the model saw.
Lookup time dropped from minutes of manual reading to seconds, and teams gained a single interface they could rely on for decision support instead of guessing which document was current.