knowhere

Overview

Knowhere is a document intelligence and retrieval infrastructure that turns complex, dirty files into persistent, navigable memory for AI agents.

It ingests unstructured documents and produces persistent, navigable memory: parsing, hierarchy reconstruction, multi-modal structuring, and graph construction in a single pipeline. Every result stays connected to its document, section, source pages, and related assets, making the output a natural fit for Agentic RAG, vector-based RAG, or any LLM workflow.

Knowhere parses files along their native hierarchy (headings, if any) and uses frontier vision models to cut the extraction errors of traditional parsers. Documents can be ingested progressively: readable and searchable first, enriched in the background. This progressive feature enables fast response in agent systems.

[!NOTE] Get started in seconds with Knowhere Cloud. Avoid the complexity of self-deployment. Use our managed API at knowhereto.ai and enjoy $5 in free credits upon registration.

📢 News

• November 2026: ⏩ Progressive ingestion. ingestion_mode=progressive publishes a readable baseline first, then enriches the rest in the background. MinerU is no longer required by default.
• November 2026: 🧰 Agent retrieval tools exposed through the API. Corpus navigation tools are exposed to built-in and external agents alike.
• November 2026: 🖼️ More image formats. .gif, .webp, .bmp, .tif/.tiff, and .heic/.heif are now supported.
• September 8, 2026: 🧭 Introducing agent-native Retrieval 2.0. Agents explore the corpus with hierarchy-aware tools (recall, grep, outline, read) and cite resolvable evidence.
• June 1, 2026: 📚 Knowhere supports ultra-long PDFs and atlases. It can process long-form PDFs (e.g., over 1000) and route technical atlases or drawing collections through a dedicated layout-aware parser.
• May 7, 2026: 🚀 Knowhere is now Open Source! We have open-sourced our entire stack for document ingestion, parsing, and agentic RAG. You can now self-host the full platform using knowhere-self-hosted. Check out our Contribution Guide to get involved!

Document Parsing: By Format, Progressively

Traditional OCR and Document Intelligence pipelines try to extract every element before a model can understand the document. On dirty PDFs and slide decks, mistakes in reading order, layout, tables, or hidden text layers can accumulate into unreliable model context.

Knowhere does not make perfect element-by-element extraction a prerequisite for retrieval. Native parsing preserves precise text and structure where they are reliable, and vision parsing uses frontier vision models to understand a page or slide as a whole, so visually complex content can still be recalled and understood without first reconstructing every element.

• Route by format: Page-based documents (e.g., PDF and PowerPoint) use vision parsing; everything else is parsed natively. Both end up in the same chunk and metadata schema, so nothing downstream depends on the source format.
• Progressive ingestion: With ingestion_mode=progressive, a document becomes readable and searchable first, then gains its outline and hierarchy, and is finally enriched in the background, with bottom-level nodes enriched on demand. Full ingestion, which publishes once parsing completes, is the default.
• Recall without brittle reconstruction: Pages can be indexed and searched through summaries, entities, source text (if any), and hierarchy even when OCR or layout extraction cannot reliably recover every component.
• One navigable memory: Text sections and vision pages become compatible hierarchy nodes with source evidence, linked assets, and cross-document relationships.

How it Works

Knowhere runs in two steps: build memory from documents, then let agents retrieve from it.

Step 1: Parse and Build Memory

• Route: Select vision parsing or native parsing according to the document format.
• Understand: Preserve native text structure where it is reliable, or understand complex pages holistically with a vision model.
• Publish: Publish a readable baseline first for progressive jobs, or publish once parsing completes for full jobs.
• Build Memory: Normalize everything into hierarchy-native chunks and store navigation trees, linked assets, citations, and cross-document relationships as agent-ready context.

Step 2: Agentic Retrieval

Knowhere provides the document-memory substrate; the agent decides how to explore it.

• Route per query: Unless a request sets use_agentic, a query router chooses classic BM25 or agent retrieval for each query.
• Provide the substrate: Knowhere exposes one corpus schema and the corpus.* tools for exact search, fuzzy recall, outlines, full reading, table queries, and assets.
• Let the agent explore: Instead of forcing every query through a fixed navigation pipeline, the agent chooses which tools to call, in what order, and how deeply to traverse.
• Stay agent-neutral: External agents call the same tools directly through MCP or POST /v2/retrieval/tools/invoke, without running Knowhere's own retrieval agent. Whichever agent explores the memory, Knowhere resolves its references into traceable documents, sections, pages, and linked assets.

FAQ

Q: What is Knowhere's relationship with MinerU?

A: PDF and PowerPoint use vision parsing: Knowhere renders source pages, combines visual interpretation with document profiling and the native hierarchy, and assembles page-grounded hierarchy nodes. MinerU is an optional fine-stage page-text extractor and is off by default. Knowhere does not treat parser-generated Markdown as the source of truth.

Q: What LLM / VLM dependencies does Knowhere have?

A: We recommend deepseek-flash as a unified model for both Text and Vision workloads. It accepts text and image input, so the same model can handle summarization, hierarchy reasoning, page understanding, and asset descriptions. Knowhere remains model-agnostic: you can use another model—or separate Text and Vision models—from OpenAI, Qwen, GLM, Volcengine, or any compatible provider.

Q: How is Agentic Retrieval different from traditional RAG?

A: Traditional RAG does a flat vector lookup and returns isolated snippets. Knowhere's agents navigate the document's section tree and cross-document graph, drilling into the most relevant regions the way a human reader would, returning traceable, well-contextualized evidence.

Q: Does it handle images and tables?

A: Yes. Knowhere extracts images and tables, runs them through VLM-assisted summarization and feature extraction, and links them back to their source section nodes. Vision parsing also retains rendered page citations, so agents can return both structured context and the visual source evidence.

Performance Benchmark

Agents using Knowhere outpe