NexusMind
A local-first agentic retrieval-augmented generation engine that combines hybrid search, a knowledge graph and a multi-agent runtime, running entirely on Ollama.
NexusMind fuses BM25 keyword search and dense vector search with reciprocal rank fusion and cross-encoder reranking, and builds an entity and relationship knowledge graph in an embedded graph database, with one-hop expansion at query time, to answer multi-hop, entity-centric questions that vector-only retrieval tends to miss.
It runs a real asynchronous multi-agent pipeline, Planner, Researcher, Critic and Writer, with typed tools, per-agent token budgets and full traceability, rather than a loop of JSON prompts. Query handling supports transforms such as hypothetical document embeddings and multi-query expansion, plus experimental semantic chunking with late chunking.
Everything runs on the user's own hardware, parsing, embedding and generation included, so private documents are never sent to a third-party API; it can be exposed as an MCP server or consume external MCP tools as agent capabilities, and offers a REST API and command-line interface for ingesting documents and asking multi-hop questions with streamed, cited answers.