knowhere

How it Works

Knowhere turns raw documents into a structured memory store that AI agents can navigate and cite. The process follows two steps:

Step 1: Parse and Build Memory

Parsing, chunking, hierarchy extraction, and graph construction are unified into one outcome: a navigable memory layer for AI agents.

• Parse: Route PDFs, Office files, images, tables, Markdown, and text to specialized parsers.
• Structure: Preserve headings, section paths, multi-modal assets, and chunk relationships.
• Build Memory: Store chunks, navigation trees, summaries, and graph links as agent-ready context.

Step 2: Agentic Retrieval

Agents retrieve by navigating memory instead of depending on a single flat vector lookup.

• Discover: Fuse keyword, path, content, and semantic signals for broad first-pass coverage.
• Navigate: Walk section trees and graph links to drill into the most relevant document regions.
• Cite Evidence: Return traceable results with source document, section, chunk, and linked image or table assets.

Performance Benchmark

Knowhere enhances the accuracy of AI agents when performing tasks (e.g., searching, modifying, and answering) in real-world data. Compared to providing raw documents directly to agents or .md/.json files produced by other parsers, Knowhere achieves higher success rates with fewer resources.

Key Advantages

• Superior Accuracy: Knowhere drastically improves both First-time Accuracy (+36% over raw docs) and Recall (+10%), ensuring agents find the right evidence faster.
• Enhanced Reliability: With user feedback, agents using Knowhere hit 79% accuracy—a significant jump compared to the ~53% ceiling of raw documents.
• Higher Efficiency: Agents require fewer loops, consume fewer tokens, and spend less time searching. By navigating a structured memory graph instead of reading monolithic texts, the token overhead drops while precision increases.

(Data generated from internal evaluation across identical agentic RAG tasks.)

[!NOTE] 📊 Benchmarks are actively expanding. The comparison above currently covers MinerU as the baseline parser. We are continuously adding more parsing tools and retrieval baselines — stay tuned for updated results.

Ecosystem

| Repository | Description | |---|---| | knowhere | This repo. Backend API and worker — document ingestion, parsing, graph construction, and retrieval. | | 🖥️ knowhere-dashboard | The web UI. Connects to the API for the full product experience. | | 🐳 knowhere-self-hosted | Docker Compose stack for self-hosted deployments. Packages the API, worker, and dashboard together. | | 🐍 knowhere-python-sdk | Official Python SDK for the Knowhere Cloud API. | | 🦕 knowhere-node-sdk | Official Node.js SDK for the Knowhere Cloud API. |

Features

• Multi-modal Parsing: High-fidelity extraction from PDF, Office, and images, preserving headings, tables, and hierarchical paths.
• Lightweight Memory Graph: Context-aware organization that links documents and chunks for better relationship understanding.
• Agentic RAG: A hybrid retrieval engine combining traditional search (RRF) with autonomous agent navigation.
• Evidence-based Citations: Every result is backed by traceable source paths, ensuring reliability for AI Agent decision-making.

Frequently Asked Questions (FAQ)

Q: Is MinerU strictly required for Knowhere to work? A: No. While MinerU is currently our default choice for parsing PDFs and PPT, because it performs the best in our experiments, any tool that can convert documents to Markdown works. Knowhere's real value lies in what happens alongside and after the initial conversion: memory-oriented parsing optimizations (fixing real-world parser deficiencies), reconstructing the hierarchical structure, normalizing multi-modal assets, and building the cross-document navigation graph.

Q: What are the LLM / VLM dependencies? A: Knowhere requires standard language models to structure the document memory. By default, it uses DeepSeek (deepseek-chat) for text/table summarization and hierarchy generation, and Qwen-VL (qwen3.5-flash) for image OCR and visual descriptions. However, it is entirely model-agnostic—you can easily configure it to use OpenAI, DashScope (Ali), Zhipu (GLM), or Volcengine (ARK) via environment variables.

Q: How does Agentic Retrieval differ from traditional RAG? A: Traditional RAG relies on flat vector similarity, which often retrieves isolated, out-of-context text snippets. Knowhere's Agentic Retrieval instead uses a multi-agent workflow to actively navigate the hierarchical section tree and cross-document graph. Agents read the document structure like a human would, drilling down into relevant sections to find precise, well-contextualized evidence.

Q: Can it handle multi-modal data like images and tables? A: Yes. Knowhere extracts inline images and tables, passes them through Vision-Language Models (VLMs) for summarization and feature extraction, and explicitly links them back to their original text chunks. This ensures that agents can retrieve and cite multi-modal assets accurately during inference.

Supported Formats

✅ Supported

• [x] .pdf .docx .pptx .xlsx .csv
• [x] .jpg .png
• [x] .md .txt .json

⏳ Coming Soon

• [ ] .epub .html .xml
• [ ] .mp4 .mp3
• [ ] .skills.md

Want to see a new format supported? Adding a parser is a great first contribution. Check out CONTRIBUTING.md to get started.

Prerequisites

• Python 3.11+
• uv
• Docker with docker compose

Quick Start

1. Sync the workspace dependencies:

uv sync --all-packages

2. Copy the environment examples:

cp apps/api/.env.example apps/api/.env
cp apps/worker/.env.example apps/worker/.env

3. Update