Skip to main content
Start by modeling the knowledge boundary before uploading documents.
  1. Create one workspace per company, customer, or tenant.
  2. Pick scopes for hard knowledge lanes.
  3. Pick projects for document collections inside those scopes.
  4. Ingest documents with consistent scope and project values.
  5. Add curated company facts for important context that is not documented yet.
  6. Search with hybrid mode while evaluating retrieval quality.
  7. Show evidence citations and curated-fact provenance in your product UI.
  8. Let users submit feedback and corrections.

​
Example model

WorkspaceScopeProjectDocuments
acmehronboardinghandbook, benefits, PTO
acmeengineeringapi-platformAPI docs, runbooks
acmebillingpricingplans, trials, refunds
acmesupportplaybooksescalation guides

​
Retrieval pattern

Use hybrid mode for most assistant experiences. Pass source chunks and extracted memories into your answer-generation layer as separate context blocks. Curated company facts should be passed as workspace-approved context and labeled separately from document citations. Use documents mode when the answer needs source wording. Use memories mode when the answer needs short extracted facts or curated company context.

​
Correction pattern

When a user flags a retrieved memory as wrong, ask for:
  • the corrected memory text
  • the reason for correction
  • whether the source chunk should also be updated
Send the correction endpoint with correctedChunkContent when the cited chunk should be updated too.