Choose recall by the evidence required

Learning objective

Diagnose vector-only recall failures and select hybrid or structured recall.

Core idea

For an architect diagnosing failed recall, returning similar passages is not the same as retrieving the required evidence. Vector retrieval ranks proximity in an embedding space. A passage can be semantically close while naming the wrong customer, version, transaction, or effective date.

Choose recall from the question’s evidence contract. Use semantic similarity for conceptual matching, hybrid retrieval when both wording and meaning matter, and structured predicates when identity or business conditions determine the answer. Diagnose which stage lost the evidence before changing the model or increasing the result count.

See 0.8 for retrieval foundations and 3.1.6 for separating retrieval evaluation from answer evaluation. This lesson concentrates on choosing the recall mechanism after a vector-only failure.

How it works

Inspect the candidate set before the generated answer

Start with a query whose expected supporting record is known. Verify that the record exists in the searchable corpus and inspect its identifier, text, version, and applicable access scope. If the source record was never indexed or its correction was not ingested, changing the ranking method cannot retrieve that missing update.

In Azure AI Search, vectorQueries specifies vector retrieval, with fields selecting the vector field and k controlling nearest-neighbor candidates. Inspect the query and returned document identifiers before diagnosing answer generation. The vector-query documentation recommends using the same embedding model for the query and the indexed documents. Check that pairing when broad retrieval quality changes after an embedding deployment change.

Vector search can return k neighbors even when those neighbors are weakly related, subject to the available documents. A populated result list therefore does not prove the index contains useful evidence. Compare retrieved candidates with the known support record, not just with a minimum result-count expectation.

Separate retrieval from later selection. The correct record might appear in candidates but be removed when context is assembled, as discussed in 2.2.3. It might reach the final prompt and still be misused. Increasing k repairs neither failure automatically and can increase noise in the context.

Select hybrid search for lexical and semantic evidence

Azure AI Search hybrid requests include both search for text retrieval and vectorQueries for vector retrieval in the same request. The two execute in parallel, and Reciprocal Rank Fusion (RRF) merges their ranked results into a unified set.

This is useful when a query contains a distinctive product code, an error string, or an exact term alongside a natural-language description. The text branch gives lexical matches another route into the candidate set, while the vector branch supports paraphrases. Test vector-only, text-only, and hybrid requests against the same known evidence to establish what the additional branch contributes.

Use the candidate-count and score distinctions in 2.2.10 when interpreting k, top, and fused or reranked results.

Add debug to the search request to inspect component subscores: "debug": "vector", "semantic" when using semantic ranking, or "all". The hybrid scoring page shows debug subscores with the stable 2026-04-01 REST API. Inspect the supporting record’s component contributions alongside the text-only and vector-only candidate lists to establish which branch retrieved it. If neither list contains it, investigate corpus coverage, query representation, or filtering before tuning the final answer. This diagnostic does not require preview query controls such as threshold.

Basic hybrid retrieval is documented with the stable 2026-04-01 REST API. The maxTextRecallSize control and vector minimum threshold are documented as preview features on their respective pages. You can diagnose and select basic hybrid retrieval without introducing either preview control. Do not make preview parameters an implicit prerequisite for the architecture.

Select structured recall for exact constraints

When the request is “retrieve the approved policy for account A and version 7,” similarity is not the acceptance criterion. Represent account, version, status, and effective time as structured metadata and use explicit predicates. Azure AI Search exposes OData filters over nonvector fields marked filterable.

A standalone filter with search=* makes the predicate the selection input; there is no lexical scoring or ranking. Text filters compare literal strings without word-breaking, and string comparisons are case-sensitive by default. A code stored as AB-42 therefore needs a deliberate normalization contract; a semantically related code is not an acceptable substitute.

In a JSON search request, filter carries the predicate. Use verified request identity and validated field values to construct it, rather than asking the model to decide which tenant restrictions are optional. select limits returned fields to the evidence the consumer needs. If the question combines exact constraints with descriptive content, apply the filter to a hybrid request: structured selection and relevance ranking are complementary.

Check the schema before depending on a predicate. The filtering documentation says an existing field cannot be changed to become filterable; add a field or rebuild the index. Writing filter in a request does not compensate for an index that lacks the required metadata or field configuration.

Structured recall also has a completeness limit. If indexed metadata is stale or missing, exact predicates faithfully select the wrong snapshot or return no match. For transactional decisions, obtain the authoritative record under the storage design in 1.2.5 rather than treating similarity-ranked text as the source of business truth.

Preserve access scope across the comparison

Use the same authorized corpus for vector-only, hybrid, and structured tests. A higher recall score obtained by removing tenant or permission constraints is not an acceptable repair.

Azure AI Search documents security filters as an application-supplied string-matching technique. Native document-level ACL and RBAC-scope enforcement is a separate preview capability. Keep those mechanisms distinct, and use the access-control design from 1.2.4. Changing the recall method does not grant additional authority to the requesting agent.

Measure whether the supporting document enters the candidates and whether the selected evidence reaches the answer. See 3.1.6 for evaluator selection. Preserve a no-evidence outcome when none of the authorized records supports the claim; a nearest neighbor is not permission to invent the missing fact.

Key distinctions and trade-offs

The following query contracts imply different starting points:

Hybrid retrieval adds candidate and merge work. Structured recall requires trustworthy metadata and explicit schema design. Vector retrieval remains useful when semantic relatedness is the intended selection criterion; the defect is using it as proof of identity or eligibility.

Architecture / decision scenario

A maintenance assistant asks for the approved repair procedure for asset P-204, revision 7. Vector-only recall returns procedures for related pump models and revision 6. The index contains revision 7, but its distinctive identifier does not enter the vector candidate set.

The team tests a hybrid request containing the code in search and the descriptive intent in vectorQueries, with "debug": "vector" enabled. Component subscores and the separate candidate-list comparison show that revision 7 enters through text retrieval. It then applies a filter over the asset identifier, revision, and approval status because the final evidence must satisfy those exact conditions. For a request containing only the exact lookup, it uses search=* with the same predicate instead of paying for similarity ranking.

All tests preserve the authorized asset scope. The team validates the selected procedure against the authoritative approval record and records a no-evidence result if it is unavailable. It evaluates candidate recall separately from the final maintenance answer, so a correct candidate ignored during generation is not misdiagnosed as a vector-search defect.

Common exam traps

• Treating k returned neighbors as proof of useful evidence.
• Treating a text filter as fuzzy or case-insensitive.
• Assuming an existing field can be made filterable in place.
• Using hybrid relevance as an exact entity or approval check.
• Assuming a filter works without filterable metadata.
• Removing access restrictions to improve recall measurements.

What to remember

Locate the missing evidence in the pipeline. Use hybrid retrieval for lexical and semantic coverage, structured predicates for exact constraints, and both when the contract requires them. Preserve authorization and distinguish candidate recall from answer correctness.

Knowledge check

References

• Microsoft, Create a vector query
• Microsoft, Create a hybrid query
• Microsoft, Hybrid search scoring
• Microsoft, Filters in keyword search
• Microsoft, Document-level access control