WeKnora knowledge bases: from documents to citation checks

A use case: ask your documents and check the source

A knowledge base retrieves relevant passages before a language model writes an answer. Product manuals, support procedures and training files are useful starting points. Fluent prose does not prove that a citation supports the answer.

The source explains Tencent’s open-source WeKnora and compares the roles of nine knowledge and memory projects. It is a documentation-based tutorial and selection guide, not a controlled performance benchmark.

Prepare the environment and two model connections

• The article uses v0.8.3, with documentation checked on October 14, 2026. This is a historical version reference, not a recommendation of the latest release.
• The Windows route uses Git, Docker Desktop and Linux containers, with available memory, disk space and dedicated ports. Locate and back up an existing installation first.
• Configure a chat model and an embedding model separately. API calls may cost money; check whether internal files may be sent to the chosen service.
• Begin with a small, readable text document. Evaluate scanned PDFs, complex tables and audio separately.

Move beyond a working login page

1. Check the selected version, configuration, containers and local health endpoint before opening the login page.
2. Create your workspace, test chat and embedding connections, then create a small test knowledge base.
3. Upload a document and inspect its parsed preview or chunks after processing finishes.
4. Ask a question supported by the file and open its citation. Then ask a question the file does not answer.
5. Record versions, backup locations and start/stop steps. Test larger collections and multiple users separately.

Acceptance questions: known and missing information

The article uses a fictional support-training file: respond within two working days after receiving fault photos, with no refund-arrival deadline. This is not an actual company policy.

A supported question

“How soon should support respond after receiving the photos?” The answer should match two working days and cite the relevant passage.

An unsupported question

“How many days until a refund arrives?” The system should acknowledge that the file does not specify this, rather than invent a deadline.

What counts as evidence

Check model connections, parsed content, the answer and its citation. A login page or successful upload alone is insufficient.

Keep purpose, costs and evidence separate

• The article discusses WeKnora, Dify, RAGFlow, AnythingLLM, GBrain, FastGPT, MaxKB, LightRAG and Open WebUI. Choose candidates by the task, then test with your documents.
• The source’s GitHub Stars are a historical attention snapshot, not accuracy, user or purchase counts.
• OCR does not guarantee chart understanding; importing a spreadsheet does not prove formula calculation. Changing embedding models may require reindexing.
• The author reviewed documentation for the official DeepSeek Harness plugin and explicitly did not install or test that integration.
• Model services, storage and maintenance have costs. Check licenses, team permissions, data handling and backups for the intended deployment.

Original article and image attribution

Source: The author’s WeKnora tutorial on WeChat ↗, by Blues, published by Blues19 Cross-border AI on October 15, 2026 at 14:40 Beijing time. This is a summary; consult the Chinese original for full steps and references.

The original Chinese diagram below explains a generic RAG process. It is not a WeKnora interface screenshot or deployment diagram.