Onboard a knowledge source without indexing everything
Onboard a knowledge source with an explicit inclusion scope, accountable owner, permission mapping, and change-and-deletion contract. Start with a representative bounded collection, inspect extracted evidence and answer quality, and expand only after lifecycle and access behavior are verified.
Inside this article 6 sections +
A connector's ability to read a repository does not mean the assistant should index the whole repository. Broad ingestion can mix approved guidance, abandoned drafts, personal notes, duplicates, and restricted customer material into one search surface. The resulting answer problems often begin before retrieval ranking or model prompting.
Source onboarding should decide what the collection is for, which questions it can support, who owns it, and how its lifecycle reaches the assistant. A bounded source contract makes those decisions reviewable and gives the connector a clear acceptance target.
Start with the questions the source should answer
Choose a collection with a recognizable purpose, such as approved product troubleshooting guides. List the question families it should support and the kinds of questions it should not answer. A troubleshooting collection may explain symptoms and recovery steps without establishing contractual entitlements.
Name a content owner who can resolve contradictions and approve scope changes. Name a technical owner for connector failures and indexing behavior. One person may fill both roles in a small team, but the responsibilities should remain distinct.
Inventory the source's structure: document types, status fields, folders, labels, versions, attachments, and access controls. Check whether the source can expose stable identifiers and incremental changes. A connector that only lists current files needs a different deletion strategy from one that provides a durable change feed.
Write the inclusion rules as data
Use a versioned ingestion contract that the connector can validate. The following synthetic example describes a deliberately narrow collection. The authorization mapping is a requirement to resolve from the source, not a claim that a label alone enforces access.
{
"collection_id": "product-troubleshooting",
"contract_version": 3,
"owner_team": "product-support",
"include": {
"root": "/approved-guides/",
"status": ["published"],
"content_types": ["text/html", "application/pdf"]
},
"exclude": {"labels": ["draft", "personal", "customer-specific"]},
"identity": ["source_id", "source_version", "tenant_id"],
"authorization": "resolve_current_source_acl",
"on_missing_acl": "quarantine",
"on_delete": "block_serving_then_remove_derivatives",
"on_parse_failure": "quarantine_with_reason"
}
Make precedence explicit when inclusion and exclusion rules overlap. In this example, an exclusion should win. Reject records missing required identity or permission information instead of placing them in a default public scope. Keep quarantined items out of retrieval until their issue is resolved.
Do not assume file paths are permanent identities. A move may change a path without creating a new document, while a reused path may contain a different document. Preserve the source system's stable identifier and a version indicator wherever available.
Sample the hard shapes before scaling volume
Verify headers, ordering, links, and extracted evidence locations.
Prove that status rules exclude the wrong source.
Check permitted users and deliberately unauthorized identities.
Trace each change through derived records and serving behavior.
Inspect extracted text alongside the original document. A PDF table may lose its column relationships; a page may include navigation text that overwhelms the useful section; an attachment may be missed entirely. Do not tune embeddings to compensate for content that was never extracted correctly.
Preserve headings, source locations, and applicable metadata through chunking. The answer interface should be able to point back to the relevant evidence. If a source cannot provide reliable locations, choose an honest citation presentation rather than inventing page numbers or offsets.
Test duplicate and superseded documents. Decide whether the connector excludes old versions, marks them historical, or preserves them for date-specific questions. A model should not have to infer organizational authority from whichever copy ranks first.
Verify access and removal before first release
Exercise the source permission mapping with synthetic users and groups. Check current membership, document-specific restrictions, and the behavior when the source's authorization service is unavailable. Index metadata is useful, but it may lag the authoritative source; define the serving check required by the sensitivity of the collection.
Change a document's permissions after indexing and repeat the same question. Inspect retrieval candidates and model context. Then delete a document while an ingestion job is running and verify that a late worker cannot publish an old version. These are onboarding tests, not optional work to defer until a user reports a leak.
Confirm how the connector discovers changes and deletions. Managed indexers and custom pipelines differ in what they detect automatically. Store a reconciliation plan for missed events, including a way to compare known source identifiers with indexed derivatives without treating a temporary listing failure as mass deletion.
Evaluate the collection as an answer source
Create questions from the collection's intended purpose and identify the supporting passages. Include scope mismatches, incomplete evidence, and questions outside the collection. Grade whether the assistant answers from the correct version and appropriately declines unsupported claims.
Separate connector acceptance from answer acceptance. A pipeline can ingest every selected file successfully while producing poor evidence because the source is ambiguous or the chunks split conditions from rules. Record the stage that needs repair rather than declaring the entire source ready based on an indexed-document count.
Define expansion gates: required fields present, permission tests passing, lifecycle behavior verified, representative extraction inspected, and acceptable answer evidence on the selected question set. Increase scope in reviewed increments so a new folder or content type does not silently inherit assumptions established for another.
Hand over an owned collection
The operating handoff should include the contract version, owners, expected update window, quarantine process, reconciliation procedure, and source-review interval. Changes to inclusion rules should be reviewed like code because they change what the assistant can know and potentially disclose.
Use freshness monitoring to operate the collection and document deletion to maintain its lifecycle. Our AI integration work helps connect a bounded source contract to dependable retrieval.