WeKnora
Import one document, ask a precise question, and inspect the citation.
WeKnora is Tencent’s open-source knowledge platform for document retrieval, RAG questions, agent workflows and generated Wiki content. It is broader than an individual skill. For a first evaluation, deploy the documented core services, configure the required models, and import one non-sensitive document with facts you already know. Check parsing and retrieval before judging the answer. A container starting successfully is not evidence that the document was indexed or that the response is supported by its citations.
01 / HOW IT WORKS
From setup to a result you can check
01
Start services
Use the documented core Compose path.
02
Configure models
Check required provider and embedding setup.
03
Import a file
Choose a small non-sensitive document.
04
Ask a question
Use an answerable fact from that file.
05
Check citation
Compare the answer with the source passage.
Suggested first-use sequence based on the official documentation. This is not a completed product test.
02 / CHOOSE YOUR TASK
Test knowledge retrieval before adding more integrations
Use WeKnora when your task is grounded in a collection of documents rather than a general model’s memory. The repository describes document, FAQ and Wiki knowledge-base workflows as well as agents and integrations. You do not need to activate every available service to evaluate a single-document question. Official overview.
Choose a source that has an unambiguous answer, such as a short internal-style test policy containing a date, a named owner and a numeric threshold. Keep it synthetic or public for the trial. Ask both an answerable question and one the document cannot support.
03 / SETUP
Start the core services and check configuration
The official quick start requires Git, Docker and Docker Compose. Its shell sequence is:
git clone https://github.com/Tencent/WeKnora.git
cd WeKnora
cp .env.example .envEdit .env for the intended environment before starting. Do not commit credentials. Review the configuration guide, especially model endpoints and embeddings. Then the documented core launch is:
docker compose pull
docker compose up -dThe README directs you to http://localhost after startup. Use your deployment’s actual host mapping if you changed it. The commands are presented for a compatible shell; on a different shell, use its normal file-copy operation for .env.
Confirm the path to an answer
- The application and required services are reachable.
- The intended model configuration works.
- The uploaded document finishes parsing.
- Retrieval finds the passage containing your known answer.
- The response cites that passage.
If parsing failed, changing the wording of the question will not repair the missing source. If retrieval succeeded but the answer is wrong, preserve the retrieved passage and response when diagnosing it.
04 / TRY ONE TASK
Use an answerable question and an unsupported question
Create a tiny test document containing: “Trial support hours are 09:00–17:00 UTC on weekdays. The escalation owner is the Example Operations Team.” After uploading it and confirming parsing, use this editorial prompt:
Use only the selected test document.
What are the trial support hours and who owns escalations?
Cite the exact supporting passage.
Then tell me whether this document specifies a weekend response time.
If it does not, state that the document does not provide that information.
Do not use web search or infer a service commitment from other sources.The first answer should match the document’s hours, timezone and owner. The second should identify the missing weekend information rather than inventing a promise. Open the cited passage and confirm that the citation points to the uploaded file, not an unrelated source.
This is a small retrieval check, not a benchmark. Before connecting real knowledge sources, also test the access permissions and update behavior required by your team.
05 / ACCESS & CONTROL
Keep deployment, data and generated content distinct
The README describes model, storage and retrieval integrations. A locally deployed application can still send content to the external providers you configure. Inspect those routes before importing sensitive files.
Generated Wiki pages and fluent answers still need source verification. Keep original documents available so you can resolve stale, unsupported or conflicting statements. A citation label alone does not prove that a passage supports the answer.
The documented stop command is docker compose down. Do not append destructive volume-removal flags as routine cleanup. Stopping services is separate from deleting data, revoking provider credentials or removing copies held by an external service.
For upgrades, the README distinguishes pulling the selected images from recreating containers. Preserve data and follow the project’s release-specific upgrade instructions; this guide has not tested a production migration.
06 / QUESTIONS
Before your first session
Is WeKnora just a skill package?
No. It is a knowledge platform with retrieval and agent workflows. Some integrations expose skills or tools, but deploying the platform is a separate task.
Why can the app open but fail to answer from my file?
Application startup, model configuration, parsing and retrieval are separate stages. Check the first failing stage instead of assuming a visible UI proves end-to-end readiness.
Does local deployment guarantee that no content leaves the host?
No. Review the actual model, parsing, storage and integration endpoints you configure. Local hosting alone does not establish their data-handling behavior.
07 / SOURCES
Sources and what was not tested
Official documentation checked on 2026-09-24. We did not install this product or run the example task for this guide. Setup commands and expected behavior are documentation-based; sample task prompts are editorial suggestions.
- Official overview and Compose quick start ↗
- Configuration and model prerequisites ↗
- Official CLI setup ↗
- Release history ↗