PolicyIQ: Hybrid RAG Assistant with Cited Answers and RAGAS Eval

This project involved building PolicyIQ, an AI assistant that answers questions about corporate policy documents with a cited, grounded answer instead of a document dump.

I built the full stack as the sole developer, and the project is open source on GitHub.

The ingestion pipeline supports PDF, DOCX, CSV and TXT files, with content-hash deduplication and parent/child chunking, processed by Redis workers.

Retrieval is hybrid: pgvector semantic search and OpenSearch BM25 keyword search are fused with Reciprocal Rank Fusion for more accurate results.

Answers are streamed on AWS Bedrock, with a groundedness check and a confidence gate that declines to answer rather than guess. Conversation memory keeps follow-up questions in context, and a Next.js chat UI provides the user experience.

The system was evaluated with RAGAS on a 12-question golden set, reaching Recall@5 of 1.0 and 94% faithfulness.

Code: github.com/Muzakiruddin566/PolicyIQ-RAG