RAG Agent
A local-first retrieval-augmented question-answering system that combines semantic and keyword search with a ReAct agent loop, citation checking, and task recovery after a restart.
RAG Agent is a document-based question-answering system where uploaded documents are indexed with Qdrant semantic search and SQLite BM25 keyword search, fused with weighted reciprocal rank fusion and an optional reranker, then handed to a ReAct-style agent loop that runs multiple tools in parallel, manages a context budget and keeps long-term memory across a conversation.
Generated answers cite their sources inline, and the system checks that citations exist and are numerically consistent before returning a streamed response, declining to answer when evidence is insufficient. It persists ingestion jobs so they can resume after a restart, and ships with authentication, backup and restore, and a one-command Docker deployment.