WeKnora: grounded answers, scoped memory and deliberate operations
Your first WeKnora knowledge base: verify parsing, retrieval and citations separately
Use a small synthetic handbook and an unanswerable question to test the path from upload to source-supported answers.
What you will learn
- Create a deliberately small test collection
- Upload completion and searchable evidence are different checks
- Make the result auditable
Before you start
- Basic HTTP and container concepts
- Understanding of documents, passages and model providers
Separate ingestion, retrieval, answer support and memory scope, then design an evidence-based acceptance exercise.
Key takeaways
- Prepare expected answers before ingestion.
- Inspect parsed chunks before blaming generation.
- Include questions that should not receive a source-backed answer.
Create a deliberately small test collection
Use three short synthetic notes: an upload limit, a versioned exception and an unrelated release note. Write expected answers before uploading. Include a question whose answer is absent, so the exercise can reveal unsupported guessing rather than only rewarding plausible text.
After service setup, log in or join an invited workspace and create a document knowledge base. Configure and test both chat and embedding connections. The quickstart says there is no default server account; registration and automatic workspace creation depend on deployment policy, while Lite has a different local onboarding path.
Upload completion and searchable evidence are different checks
The documented parsing states progress through pending, processing, finalizing and completed. Wait for completion and inspect the resulting chunks before asking the acceptance questions. A successful HTTP upload only proves receipt, not that the expected passage became searchable.
Use knowledge-search when you want retrieval results without composed answers. Then compare Quick Q&A on the same knowledge base. If the expected passage is missing, inspect parsing, chunking and retrieval; if it is present but the answer misstates it, focus on context assembly and generation.
Make the result auditable
Open each citation and compare the answer with the exact source version. Test the absent-answer question as well as direct and paraphrased questions. Record unexpected answers as failures or unresolved observations, not as successes because the interface streamed smoothly.
The sample worksheet is a plan, not a fabricated upload transcript. Do not copy the documentation’s demonstration password or full-access key into production. Start with the minimum permissions needed for this collection and keep unrelated documents outside the exercise.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Create three synthetic, versioned notes.
- 2
Test chat and embedding connections.
- 3
Wait for completed parsing and inspect chunks.
- 4
Compare retrieval-only results, answers and citations.
Copy-ready example
{
"exercise": true,
"questions": [
{
"query": "What is the upload limit?",
"expectedSource": "policy-v1"
},
{
"query": "What discount is offered?",
"expectedSource": null
}
],
"parseCompleted": null,
"retrievalChecked": false,
"citationsChecked": false
}Frequently asked questions
Does an upload response prove the document can answer questions?
No. Parsing and indexing still need to complete.
Why ask a question absent from the documents?
It tests whether the system signals missing evidence rather than inventing support.
Sources
- WeKnora / website-docs/01-getting-started/03-quickstart.mdSource checked 2026-09-14
- WeKnora / website-docs/02-architecture/04-rag-pipeline.mdSource checked 2026-09-14