Loading...
Loading...
A production-grade Retrieval-Augmented Generation system that analyses financial reports with GPT-4o, achieving 0.883 overall RAGAs score with perfect context recall, featuring semantic reranking, confidence guards, and real-time streaming.
FinSight is a production-ready RAG system built for financial document analysis. Users upload annual reports, 10-K filings, or earnings releases and ask natural language questions — the system retrieves the most relevant passages, reranks them using a Cohere cross-encoder, and streams cited answers token by token via GPT-4o. Built on a six-layer pipeline covering ingestion, chunking, embedding, retrieval, reranking, and generation, the system was rigorously evaluated using the RAGAs framework achieving 0.992 answer relevancy and 1.000 context recall. A confidence guard refuses to generate when retrieval scores fall below threshold, prioritising accuracy over coverage — critical in financial contexts where a wrong answer is worse than no answer.



