Tencent Open‑Sources WeKnora: An AI‑Powered Document Understanding Framework

Developers often waste time on massive unstructured documentation, and enterprises struggle to build private knowledge bases that are secure.

WeKnora is Tencent's newly open‑source, one‑stop framework for document understanding and semantic retrieval, built on the self‑developed IMA kernel.

Core capabilities

It ingests unstructured files such as PDF, Word, and scanned images, performs intelligent parsing and structural conversion, and enables accurate question‑answer interaction.

The framework tightly integrates large language models (LLM) with retrieval‑augmented generation (RAG) to address three common pain points: slow parsing speed, broken layout after conversion, and weak semantic understanding.

Advantages over existing solutions

Top‑level parsing : Powered by a Chromium‑optimized IMA rendering engine, it handles complex mixed‑media layouts and OCR‑extracted text. Official tests show parsing efficiency more than 300% higher than traditional tools.

Flexible deployment : Supports local, Docker‑one‑click, and private‑cloud deployments, meeting strict data‑security requirements for finance, legal, etc., with full‑stack monitoring.

Modular, low‑code design : Visual drag‑and‑drop UI lets non‑technical staff build knowledge bases, while developers can customize config files, choose vector databases (Milvus, Chroma), and switch LLM providers (Qwen, DeepSeek, Ollama).

WeChat ecosystem integration : As the core framework of the WeChat Open Platform, it can be deployed to official accounts or mini‑programs without extensive coding, enabling fast intelligent customer service.

Architecture

Core engine : Chromium‑based IMA rendering kernel for stable parsing of complex documents.

Language stack : Go for backend processing and Vue for frontend interaction.

AI algorithm stack : Supports over 20 major LLMs, uses sliding‑window and parent‑child chunking, combines BM25 keyword search with dense vector retrieval; top‑10 retrieval precision reaches 89%.

Infrastructure : Compatible with Milvus and Chroma vector stores, deployable via Docker or Kubernetes, with Nginx reverse‑proxy support.

Additional features

Supports more than ten document formats (PDF, Word, Excel, PPT, Markdown) and automatic web page crawling.

Integrates with Feishu, Notion, Yuque for full‑sync of knowledge sources.

Layered knowledge modeling automatically splits long documents while preserving semantics.

Agent‑driven Wiki mode generates standardized Markdown encyclopedia pages and visual knowledge graphs.

RAG‑based intelligent Q&A marks source citations to avoid hallucinations; includes ReAct Agent for complex task decomposition and a Data Analyst Agent for data‑analysis scenarios.

Web management console, Chrome bookmark plugin, and mobile mini‑program for anytime knowledge‑base management; public API enables integration with Enterprise WeChat, Slack, etc.

Deployment guide

Supported OS: Linux (Ubuntu 20.04+ recommended), Windows (with WSL2), macOS. Prerequisites: Docker, Docker‑Compose, Git.

After cloning the repository, run the one‑click script: bash docker/start.sh The system pulls images, configures the environment, and starts the service. Access the UI at http://localhost:8080 .

Images illustrating the UI and architecture are included above.