Retrieval must enforce a userβs permissions before results enter the prompt. A relevance filter can narrow a candidate set, but it becomes an access control only when a trusted layer derives and applies it from verified identity. The buyer needs to identify the component that owns that decision.
My take: evaluate Vectaraβs authorization and filtering model for a managed retrieval stack. Compare Ragie when your pipeline is organized around connector ingestion and document metadata. In both cases, prove what the platform enforces itself and what your application must supply.
| Candidate | Documented building block | Confirm in the trial |
|---|---|---|
| Vectara | Attribute-based access control and metadata filters | Which identity is evaluated and how policies apply across corpora and query paths. |
| Ragie | Metadata filters constrain retrieval | Can a server enforce group filters on every query, and how are ACL changes and caches invalidated? |
Make revocation the central test
Index two documents containing the same terms. Give one to both teams and restrict the other to finance. Query as a permitted and denied user, including a direct fetch and a cached answer. Remove finance membership while the permitted user still has a session. Repeat retrieval, reranking, citation opening and any document preview. Define the maximum interval before revocation must take effect.
Treat snippets, stored conversations and exports as separate surfaces. A correct filter on semantic search does not prove the citation link or tracing interface applies the same policy. Check default behavior when group synchronization fails or the document has no ACL metadata. Use only synthetic data during the trial.
Price the boundary you can operate
Ask which identity integrations, data partitions, metadata operations and retention controls belong to the quoted plan. Include the cost of updating access rules and deleting stale indexed copies. Keep authorization in your existing application if it can apply an explicit policy at every retrieval path. Buy managed controls when the vendor can demonstrate enforcement that reduces code and operational work without weakening revocation.
Vectara for managed retrieval boundaries
Start with the identity and authorization documentation. Draw the path from the authenticated person to the query request. Determine which access policy is applied at each layer, how corpora are selected and who has permission to update filters. The exact policy behavior depends on the feature and configuration you buy.
Ask the supplier to demonstrate two corpora with different permissions and two documents inside a shared corpus with different labels. Inspect the results from the same identity in both configurations. Decide whether the smaller permission boundary can be enforced without growing the number of corpora every time a new group is added.
Ragie for ingestion and application-supplied filters
Evaluate a trusted backend that derives a filter from a verified session. Ask how connector metadata arrives, which fields are stored per document and what happens when the ACL data changes. The documented retrieval filter is a building block. Confirm the entire authorization path before enabling a client that can issue queries.
Try a request with a deliberately absent or forged filter in an authorized sandbox. The server should supply the permitted restriction itself or deny the request. Check whether a debugging endpoint or direct document lookup bypasses that same logic. Retain a trace for the decision without logging the restricted chunks.
Account for the document lifecycle
Consider a document split into several chunks. One version may be reindexed while older chunks are still present. When an ACL changes, test whether the old chunks retain permission to appear in search, summaries or generated citations. Track source version, document ID and policy revision separately. Define which job makes the index converge after an interrupted update.
Use independent test identities for the content owner and the person querying. Test a link copied from a permitted answer to a denied browser session. Check that the target applies its own access policy. Re-test after group removal, with the browser cache warm, so the proof covers the entire reader workflow.
Questions to settle before choosing
Can I trust a filter supplied by the browser?
Derive access decisions on a trusted server from a verified identity. Let users request ordinary search constraints only within that scope.
Do separate corpora solve every permission problem?
They can make boundaries simpler when they align with real tenants. Shared corpora or within-tenant document restrictions still need a policy.
What if group synchronization is temporarily unavailable?
Define how stale group state is handled. Verify fail-closed behavior where the system cannot determine access.
Does hiding a citation protect the document?
Protect retrieval, preview and download separately. A hidden link cannot prevent exposure if a server sends unauthorized text into the prompt.
Include membership changes in the access model
Many demonstrations give each test user a fixed set of groups. Add a guest, someone who belongs to two business units and a user who changes teams during the trial. Document where user and group identifiers originate and which service updates membership. If the repository and application represent groups differently, require an explicit mapping. Test how the platform handles a missing mapping and what administrators see when access is uncertain.
Trace the rule from the document through indexing and retrieval. A collection-wide rule may work for an employee handbook and be too broad for a client folder with per-project sharing. An application-side filter may be useful when it runs before candidates reach the prompt and logs. Ask which fields are visible to the retrieval service and how the same policy applies to a citation preview or a cached answer.
Prepare for an incident
Choose one test document and move it from an internal folder to a restricted folder. Capture when the source change occurred and when the search service stopped returning it. Test a previously saved response as well as a fresh question. If a cached citation still opens in another system, determine which service owns that exposure. Verify that the rollback of an application release does not also roll back an access change. Ask how to disable a connector quickly, who can authorize that step and what evidence the provider retains for post-incident analysis.
Sources checked 12 November 2026. The comparisons are based on published documentation. The trial scenarios are proposed acceptance tests.