Build compatible representations
An embedding model maps an input to a vector intended to capture relationships useful for a task. For retrieval, queries and documents may use different task types from the same supported model; follow that model’s documentation rather than assuming every task type is interchangeable. Keep model version, task type, dimension, normalization, and document-processing version with the index. If the embedding model changes, assess compatibility and plan a controlled re-embedding and index migration.
A search system embeds a user question and its policy documents. Explain why recording the query and document embedding configurations matters when the team changes the embedding model.
Choose similarity and indexing deliberately
A distance or similarity measure ranks candidate vectors; choose one compatible with the embedding model and normalization. Exact search compares against the full eligible collection, while an approximate nearest-neighbor index trades some exact-neighbor recall for performance. Apply authorization and metadata filters to the eligible corpus and evaluate their interaction with search. A nearest vector is not automatically relevant, current, or authorized evidence.
An approximate search returns 8 of the 10 neighbors returned by exact search for the same query and eligible corpus. Index recall at 10 is 80%. Explain why this does not tell you whether any of the ten passages actually answer the user’s question.
Evaluate retrieval in two layers
First compare approximate results with exact neighbors using the same vectors, distance, filters, and top-k to quantify the index tradeoff. Separately use human-judged queries and relevant documents to measure whether useful evidence is retrieved and ranked. Inspect missing documents, chunk boundaries, language coverage, stale sources, and access control. Evaluate the final answer for correctness and support after retrieval; better nearest-neighbor recall alone does not prove better answers.
The index has 98% recall against exact search, but users still get irrelevant passages. List three explanations unrelated to approximate-index accuracy and propose one test for each.
Operate the index as a data product
Track source updates, deletions, permissions, embedding versions, indexing lag, query latency, and cost. Test a new index alongside the current one using representative queries before switching traffic, and retain a rollback path. BigQuery supports vector search with an index for approximate search or brute-force comparison; other serving products have different operational and access characteristics. Verify current documentation and deprecation notices before choosing an implementation.
A document is deleted or a user loses access. Design a test that confirms the old content is no longer retrievable by that user, including any synchronization delay and cache behavior.
An approximate vector index returns eight of the ten exact nearest neighbors. What does its 80% index recall establish?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.