When preparing for Databricks GenAI Engineer Associate, Databricks renamed Vector Search to Databricks AI Search in 2026. The queued term “Mosaic AI Vector Search” still describes the technology many teams built against, but current documentation uses AI Search for the managed retrieval service. The rename also reflects a broader product direction: the service is not limited to pure nearest-neighbor vector lookup. It supports semantic retrieval, hybrid search, metadata-aware filtering, and newer full-text capabilities inside a governed Databricks data platform.
For Databricks GenAI, AI Search is the retrieval layer that turns governed Delta data into low-latency search indexes for RAG, recommendation, and similarity workloads. The engineering challenge is not merely creating an index. It is deciding what data belongs in the index, how it stays synchronized, how queries are transformed, how ranking is evaluated, and how access controls remain aligned with the source data.
AI Search indexes are derived serving structures, not the source of truth
An AI Search index is built from or populated with data that originates elsewhere. For Delta Sync indexes, Databricks incrementally synchronizes a Delta table into the search structure. For direct vector access, applications write vectors and metadata through the API or SDK. In both cases, the index exists to serve retrieval efficiently; it should not become the only place where authoritative document state lives.
Keeping a governed Delta table as the source of truth improves reproducibility. Teams can inspect the exact text, metadata, document version, and permissions used to build an index and can rebuild the index when embedding models, chunking strategies, or schemas change. Treating an opaque retrieval index as primary storage makes those migrations much harder.
Delta Sync and direct vector access serve different ingestion models
A Delta Sync index is attractive when source records already live in a Delta table and should flow into retrieval automatically. Databricks manages incremental synchronization, and standard endpoints can use Change Data Feed to track updates. That fits document pipelines where parsing and chunking land governed records in Delta before indexing.
Direct vector access gives the application explicit control over inserts and updates. It can be useful when vectors are produced outside a Delta-centered pipeline or when the application needs a custom lifecycle. The cost is that the application now owns synchronization correctness. If a document changes and the index is not updated, retrieval can silently return stale evidence.
Embeddings are only one part of retrieval quality
A common mistake is to blame the embedding model for every bad search result. The deeper retrieval analysis in Databricks AI Search begins earlier: document parsing, chunk boundaries, duplicated boilerplate, metadata design, language normalization, and source quality all determine what the index can retrieve.
Query-time choices matter too. Pure semantic similarity may retrieve conceptually related but operationally wrong documents. Keyword evidence can be crucial for error codes, product names, identifiers, or legal clauses. Hybrid search combines lexical and semantic signals so exact terms can influence results without giving up conceptual matching.
Metadata filters are part of relevance and security design
Metadata such as product, region, document type, customer, language, effective date, and access classification can narrow the candidate set before or during retrieval. That can improve both relevance and governance. A support agent answering a question about one product line should not rank documents from unrelated products just because their language is semantically similar.
Metadata design should align with catalog and application identity. The broader Unity Catalog model provides the governed data boundary; retrieval metadata provides application-level structure inside that boundary. Do not use metadata filters as a substitute for real permissions when data must be access-controlled. A filter passed by the client is not a security boundary unless the service independently enforces the user’s entitlement.
Chunking decisions often dominate index behavior
The production RAG stack usually converts documents into chunks before indexing. Chunks that are too large can mix several topics and dilute similarity. Chunks that are too small can lose context and force the generator to reconstruct relationships from fragments. Repeating navigation text or legal boilerplate in every chunk can swamp the embedding space with non-informative content.
Evaluate chunking with real questions. A useful chunk should contain enough context to answer or support a specific user intent while remaining focused enough to rank well. Tables, code, transcripts, and policy documents may need different parsing strategies. One global character-count splitter is rarely optimal for an enterprise corpus.
Index freshness needs an explicit service objective
A retrieval system can be accurate and still be operationally wrong if it serves stale information. Define how quickly source changes must appear in the index. Product documentation may tolerate minutes of lag; incident-response knowledge or rapidly changing inventory may require much tighter synchronization. Delta Sync modes and pipeline schedules should be selected against that objective.
Freshness monitoring should compare source and index state, not just report that an index endpoint is healthy. If the pipeline that creates chunks stops updating, the endpoint can continue returning results from yesterday. Observability needs to cover ingestion time, last successful sync, failed records, embedding errors, and query health.
Reranking and query transformation can improve difficult retrieval
The first retrieval pass does not have to be the final evidence set. Applications can rewrite ambiguous queries, generate multiple search formulations, combine semantic and lexical retrieval, and rerank the candidates using stronger models or domain-specific scoring. Those techniques can materially improve recall and ordering for complex questions.
Every additional step adds latency and cost, so improvements should be measured. A reranker that raises answer quality by two percentage points but adds two seconds to every request may be unacceptable for an interactive assistant. Offline retrieval evaluation and trace analysis help quantify whether the extra layer is worth it.
Retrieval evaluation should use labeled information needs
Search quality cannot be judged reliably by reading a few top results. Build an evaluation set with user questions, known-relevant documents or passages, and important hard negatives. Measure whether relevant evidence appears in the candidate set, where it ranks, and whether filters accidentally remove it. Metrics such as recall at K and ranking-oriented measures provide a more stable baseline than subjective demos.
Then connect retrieval evaluation to generation evaluation through MLflow for GenAI. A bad final answer may originate in retrieval, prompt assembly, or the language model. When traces preserve the retrieved candidates and scores, teams can assign the failure to the right layer instead of repeatedly rewriting prompts.
The current product goes beyond ‘vector’ in both name and capability
The AI Search rename is useful because modern retrieval increasingly combines vector similarity with keyword and structured signals. Databricks documentation now describes dedicated full-text options in addition to vector and hybrid search modes. The service is evolving into a general retrieval platform rather than a narrowly branded vector database feature.
That does not make vector embeddings unimportant. They remain central for semantic similarity. It means architects should select the query mode that matches the information need. Exact codes and names often benefit from lexical matching; conceptual questions benefit from embeddings; many enterprise queries need both.
A production index needs governance, evaluation, and lifecycle management
Creating an endpoint and index is only the beginning. Production teams need source ownership, schema-change handling, embedding-model versioning, rebuild procedures, freshness monitoring, query evaluation, cost tracking, and access review. Those controls keep Databricks AI Search aligned with the data and applications that depend on it.
The most durable design treats retrieval as a product with measurable quality. The source table is governed and reproducible, the index is refreshed to a known service level, search behavior is evaluated on real questions, and downstream agents consume the results through a stable interface. The Vector Search name may be legacy, but the engineering discipline behind reliable retrieval is more important than the label.
Permissions around the index must match the application boundary
AI Search resources require catalog and index privileges, and Databricks Apps can be granted access through service principals. The platform team should avoid giving an application broader catalog privileges than its retrieval task requires. If an agent only needs one index, its identity should not automatically receive wide table access simply because the index was built from those tables.
At the same time, index permissions must be reviewed when source governance changes. If a source table becomes more restricted but an existing index remains broadly callable, the retrieval layer can become an unintended bypass. Treat source, index, and application permissions as one dependency chain and include all three in access reviews.
Failure analysis should distinguish empty results, stale results, and bad ranking
Search incidents often look identical to the user—“the agent could not find the answer”—but the causes differ. An empty result can come from filters that exclude every document, a synchronization gap, or an embedding/query mismatch. A stale result suggests freshness failure. A relevant document at rank 30 indicates ranking or query formulation problems. Each category needs a different fix.
Instrument retrieval with enough metadata to classify the failure. Record query mode, filters, index version or sync state, candidate identifiers, scores, and latency in traces where privacy permits. That evidence makes it possible to improve the retrieval system systematically instead of repeatedly adding prompt instructions such as “search more carefully.”