Knowhere

Knowhere is a document parsing and retrieval system that turns complex, messy files into persistent, navigable memory for AI agents, including local and offline document collections. One pipeline parses each document, reconstructs its hierarchy, structures images and tables, and builds a graph, and every result stays linked to its document, section, source pages, and related assets. The output feeds agentic RAG, vector-based RAG, or other LLM workflows.

Parsing runs on two tracks that produce the same chunk and metadata schema. The Text Track keeps precise extracted text and native structure where it is reliable. The Vision Track sends PDF and PowerPoint pages to a vision model that reads each page or slide as a whole, so visually complex content stays recallable without reconstructing every element through OCR. Long PDFs with hundreds of pages and technical atlases go through a layout-aware parser, and MinerU remains the raw PDF extractor for the older chunk-based pipeline.

Retrieval gives the agent tools instead of a fixed pipeline: document outlines, structural filters, exact search, fuzzy recall, full reading, assets, and cross-document relationships. The agent chooses which tools to call and how deeply to traverse, and Knowhere resolves its references into documents, sections, page numbers, and rendered page images as citations. The same corpus serves built-in agents and external agents over MCP, and classic top-K retrieval remains available.

In internal evaluations on agentic RAG tasks, agents working from Knowhere output scored 36% higher first-try accuracy and 11% higher recall than agents working from raw documents, and reached 79% accuracy with feedback against roughly 53% on raw documents. The backend API and worker are self-hostable with a separate web dashboard, and a managed cloud API has Python and Node.js SDKs.