WeKnora: grounded answers, scoped memory and deliberate operations
Build a WeKnora evidence workbook: missing passages, memory gaps and answer support
Design a read-only learning project that links source versions, recall behavior and claim-level acceptance without importing private conversations.
What you will learn
- Make missing evidence a first-class result
- Add a memory coverage exercise
- Prefer a useful evidence table over decorative motion
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
- Record unsupported and obsolete answers explicitly.
- Display candidate coverage separately from final output limits.
- A synthetic workbook can teach the mechanism without private data.
Make missing evidence a first-class result
Create a workbook of synthetic questions, expected source IDs, retrieved passages, answer claims and unresolved checks. Include an absent-answer case and an obsolete-policy case. The useful output is not a universal green score but a precise record of which claim lacks support.
Keep the question set versioned alongside the synthetic documents. When chunks, models or retrieval settings change, compare the same cases. A passing result on yesterday’s source version cannot certify today’s changed policy without another check.
Add a memory coverage exercise
Reuse the low-importance relevant-memory example from the source chapter. Show the old importance-first window beside scope-first relevance selection, then display candidate growth and final budget separately. Scope decoys and expired items should remain excluded even when they have the highest assigned score.
The twelve model tests are a teaching aid for this workbook, not evidence that the actual service passed an authorization audit. Real memory data, model embeddings and database migrations remain out of scope until a separate controlled validation is authorized and configured.
Prefer a useful evidence table over decorative motion
A two-lane comparison or a compact graph can make candidate exclusion visible. Three-dimensional animation would be justified only if it explains a genuinely spatial relationship, which this ranking exercise does not require. Keep labels readable and provide text equivalents for each state.
Start read-only with synthetic records. Automated ingestion, scheduled rechecks, sandbox execution and publication are later features with separate authority and retention needs. This proposal creates none of those services; it gives learners a concrete way to inspect evidence before trusting a knowledge answer.
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
Version the small corpus and expected answers.
- 2
Record passages and claim-level support.
- 3
Add scoped memory counterexamples.
- 4
Review the table before considering live automation.
Copy-ready example
{
"proposal": true,
"corpusRevision": "synthetic-v1",
"expectedSource": "policy-v1",
"retrievedSource": null,
"claimSupported": null,
"scopeDecoysExcluded": null,
"privateMemoryImported": false,
"scheduledJobCreated": false
}Frequently asked questions
Is this workbook an existing WeKnora feature?
No. It is an editorial learning-project proposal.
Does the proposal connect to private memories or schedule jobs?
No. It uses synthetic records and creates no external service or schedule.
Sources
- WeKnora / website-docs/01-getting-started/03-quickstart.mdSource checked 2026-09-14
- WeKnora / internal/application/service/memory/recall_trace.goSource checked 2026-09-14
- WeKnora / internal/application/repository/memory_vector.goSource checked 2026-09-14