Module assessment

1.

A retail team enables full-text search on a product container and indexes the description property. Every product's description holds a generated sentence of the form The product called "<product name>". Searches for helmet match the expected products, but searches for category words absent from product names don't find them. The team wants one indexed text property to cover both names and categories. What's the appropriate response?

Add the remaining product properties to the full-text policy so that every path is searchable.

Build a property that concatenates the words shoppers use, such as the product name and category, and index that property instead.

Switch the queries from FullTextContains to CONTAINS so that matching works on the raw characters.

2.

An application stores 1,536-dimension embeddings for 2,000,000 documents in one container, partitioned by tenant. Every search query filters to a single tenant, and the largest tenant holds about 20,000 documents. Which vector index type fits this workload?

quantizedFlat, because the filter narrows each search to roughly 20,000 vectors.

flat, because it gives 100 percent recall and the per-query scope is small.

diskANN, because the container holds far more than 50,000 vectors in total.

3.

A developer writes SELECT TOP 10 c.name, FullTextScore(c.searchText, "mountain") AS relevance FROM c ORDER BY RANK FullTextScore(c.searchText, "mountain") and the query is rejected. What explains the failure?

The same system function can't appear twice in one query, so the projection has to be removed and replaced by a second query.

FullTextScore is valid only inside an ORDER BY RANK clause or as an argument to RRF, so it can't be projected in the SELECT list.

The container is missing a full-text index on /searchText, and the score can only be projected once the index exists.

4.

A change feed consumer regenerates an embedding for each change it reads and writes back the updated item to the same container. After one catalog update, it keeps running indefinitely, calling the embedding model continuously even though nothing else writes to the container. Checkpoints advance successfully, and each cycle reads a newer version of the item. What's happening?

The lease container is misconfigured, so the processor never records a checkpoint and reprocesses the same changes.

Latest-version mode replays the full history of each item, so a single update produces one change per previous version.

The consumer's own write is a change, so it reappears in the feed the consumer is reading and triggers another refresh.

Check your knowledge

Feedback