# CausalGraph AI

An open-source application that turns long-form ESG corporate disclosures into searchable, cited evidence through retrieval-augmented generation and an optional knowledge graph.

CausalGraph AI ingests corporate reports as PDFs, DOCX files or plain text, converts them into cleaned overlapping chunks, and answers questions from retrieved evidence rather than unconstrained model memory. Responses stream incrementally and can include retrieved chunk references, source metadata and graph-backed evidence when Neo4j is configured.

Retrieval is configurable: vector search, BM25 fusion, multi-query expansion, HyDE, reranking, graph context and query decomposition are all controlled through environment settings. Redis-backed sessions preserve short-term chat state and recent messages, and entity-aware document scoping keeps follow-up questions from retrieving evidence for the wrong company or report.

The stack combines a React frontend with a FastAPI backend, and defaults to a DeepSeek model over an OpenAI-compatible API, with OpenAI or Anthropic available as an explicit rollback. According to its README, the codebase is being rebuilt into a legal contract-review agent for in-house counsel.

Website: https://casualgraphai.vercel.app