Knowhere
A document parsing and retrieval system that turns unstructured files into persistent, navigable memory for AI agents and retrieval-augmented generation.
Knowhere ingests unstructured documents, such as dirty PDFs and slide decks, and turns them into persistent, navigable memory for AI agents. A single pipeline handles parsing, hierarchy reconstruction, multi-modal structuring and graph construction, and every extracted piece stays linked back to its source document, section, page and related assets.
Version 2.0 adds a vision track alongside the original text track: text-native documents keep their precisely extracted structure, while complex PDFs and slide decks can be understood directly as page images by vision models, and both tracks converge on the same memory schema, hierarchy and citation model. The output is designed to plug into agentic RAG, vector-based RAG or other LLM workflows, and the project can also be self-hosted.
It is written in Python and released under the Apache 2.0 license.