finsight-rag

Financial Research Assistant (Equity Research Agent) is an intelligent Retrieval-Augmented Generation (RAG) assistant designed for equity research, FP&A workflows, and financial document analysis. Users can drop in financial reports such as 10-Ks, 10-Qs, earnings call transcripts, or equity research notes as PDFs and immediately ask natural language questions about valuation assumptions, revenue drivers, capital expenditures, margins, and guidance. The system combines an embedded vector store and compute orchestrator (Pixeltable) with local vector embeddings (intfloat/e5-large-v2) and ultra-fast LLM inference hosted on Groq (openai/gpt-oss-120b). All answers are strictly grounded in retrieved document chunks with zero hallucination. The agent strictly answers from retrieved chunks and alerts the user if the document does not contain sufficient information. Pixeltable views automatically split uploaded PDFs into 300-token chunks and maintain vector embedding indexes declaratively. The system runs against an embedded PostgreSQL instance (pixeltable_pgserver) — no external hosted vector database subscription required. Embeddings run locally via sentence-transformers, caching weights so subsequent runs require no Hugging Face network requests. A secondary vector store export to LanceDB is demonstrated in an exploratory notebook.