IlyaRice/RAG-Challenge-2
A winning RAG solution that combines vector search, LLM reranking, and chain-of-thought prompting to answer questions about company annual reports.
Velocity · 7d
+5.3
★ / day
Trend
→steady
star history
This repository contains the winning implementation for the Enterprise RAG Challenge 2, covering two prize categories. The system uses custom PDF parsing with Docling, parent document retrieval via vector search, LLM-based reranking for context relevance, structured output prompting with chain-of-thought reasoning, and query routing for multi-company comparisons. It combines OpenAI and Gemini models for generation, and includes test datasets from the competition.