RAG system for an internal knowledge base

The challenge

Hundreds of policy documents, manuals and reports across different formats. Employees spent an average of 45 minutes per search. The existing search only worked on exact keywords.

The approach

A Retrieval-Augmented Generation (RAG) system that:

1. Processes documents and stores them as searchable embeddings
2. Accepts questions in natural language
3. Finds relevant passages via semantic search
4. Formulates a coherent answer with source references

Architecture

• Document processing: PDF/Word extraction, chunking with overlap
• Embedding: text-to-vector conversion for semantic comparison
• Vector store: efficient storage and search
• Query engine: question → retrieval → LLM → answer with sources
• Frontend: a simple chat interface for end users

The result

• Search time: from 45 minutes to 30 seconds
• Accuracy: 92% correct on validation
• Adoption: 80% daily use within 2 weeks

Technology

Python, LangChain, OpenAI embeddings, ChromaDB, Streamlit