Internship · OMFYS Technologies · 2025
A way to ask questions about long policy documents at OMFYS and get the answer from the right section, instead of reading the whole document.
- 60%
- more accurate answers, tested with real clients
01The problem
Policy documents are long and dense, and the answer to a specific question is usually one clause buried somewhere in the middle. A basic RAG setup cuts documents into fixed-size pieces and pulls back whichever ones look most similar to the question, which often means the wrong piece, or half of the right one.
02How I worked
I didn't build this on my own at a desk. I sat down with my mentors and the teams involved and went through the whole lifecycle of the system with them, end to end, from what it needed to do to how it would work for clients.
03What I built
A RAG system for policy documents that uses a knowledge graph alongside semantic search, so it can follow how sections and terms connect instead of only matching similar words.
Adaptive chunking, which splits each document based on its content instead of into fixed-size pieces.
Graph-aware retrieval, which uses the knowledge graph to pick the passages that actually answer the question.
Redis caching and Hugging Face embeddings, so repeat questions don't redo the same work.
04How it went
We tested it with real clients, and answers were 60% more accurate.
