Knowledge & actions · Explore this field ↗ · Implementation · 2 min read
Ground answers in knowledge users can inspect
Retrieval is a publishing and operations problem before it is a model feature.
The corpus is the product
Retrieval quality cannot exceed source quality. Inventory owners, effective dates, audiences, jurisdictions, and supersession rules before embedding documents. Remove navigation clutter and duplicated boilerplate. Split by coherent decisions rather than arbitrary token length. Preserve source identifiers so an answer can link to the exact policy or paragraph.
Permission travels with content
A user must not retrieve a document merely because the model can vector-match it. Apply authorization before retrieval and again before tool use. Separate public help, account-specific records, employee material, and restricted operational notes. Log the documents considered without exposing their contents to unauthorized analytics.
Absence needs a designed response
When retrieval finds nothing strong, do not ask the model to fill the gap from general knowledge unless the product explicitly permits it. State that the approved material did not answer the question, collect clarifying context, or hand off. Google’s data-store tools expose source and fallback controls; the broader design lesson is to make empty retrieval a first-class state.
Operator note
Run a freshness report and an entitlement test before relevance tuning. Evaluate citation correctness separately from answer correctness. A cited page can be real yet irrelevant, outdated, or misapplied.
The source maintenance loop
Assign each collection an owner and review interval. On every change, record effective date, superseded page, audience, and jurisdiction. Re-index only after validation, then run questions whose answers should change and questions that must not change. Keep a quarantine for drafts and expired material instead of deleting provenance. When operators flag a bad answer, diagnose retrieval, source content, ranking, instruction, or generation separately. Updating a prompt cannot repair a policy page that is itself ambiguous.
Freshness signal
Show the source title and effective date when policy currency matters. If several documents conflict, stop and route the conflict to the content owner instead of asking the model to choose. That disagreement is an operational defect, not a conversational opportunity.
Primary reading
Sources and limits
These links support the architecture, policy, or product behavior discussed above. Vendor documentation describes vendor features; it is not independent proof of performance. Current details should be rechecked before a production decision.