Knowledge · Guide
Product codes need more than semantic similarity
A retrieval test that gives exact identifiers and ordinary language different jobs.
Distinguish two kinds of intent
A customer asking for a particular part number is not asking for something that merely sounds similar. Microsoft’s hybrid-search documentation describes combining text and vector retrieval. For a small distributor, the editorial recommendation is to make exact identifiers a first-class test case. Preserve punctuation, leading zeros, and supplier aliases. Decide whether a near match should be shown as a suggestion rather than treated as an exact answer.
Build mixed-language examples
Use illustrative queries such as a Chinese description plus an English model code, a misspelled product name, and a correct code with an obsolete brochure. Separate identifier correctness from descriptive relevance. A result may match the code but be the wrong version. Include both the approved document and a plausible distractor in the test corpus so the system must demonstrate the distinction rather than succeed by default.
Workflow at a glance
Keep filtering outside the guess
Apply business filters such as approved status, customer scope, and valid date before the final answer is assembled. The language model should not be the sole guardian of those boundaries. Show users the identifier and version in the result card. When no exact match exists, offer a controlled next step: ask for a photograph, search a supplier catalogue, or send the issue to staff. Avoid silently substituting a similar item.
Evaluate the handoff
Measure whether staff select the correct item and how much verification is required. Retrieval ranking is a diagnostic, not the complete outcome. Keep failed queries as regression cases whenever a new catalogue or naming convention is added. A hybrid design is not automatically better; it earns its place only when a controlled comparison shows that exact and descriptive questions are handled more reliably at acceptable effort and cost.
An exact-match challenge
Prepare similar identifiers that differ by one character, together with obsolete and current references. Ask which cases should produce an exact match, a suggestion, or no answer. These distinctions should be visible in the interface rather than buried in a relevance score.
Sources and further reading
- Hybrid search using vectors and full-text searchHybrid search using vectors and full-text searchMicrosoft · Documentation ·
- Create an index for multiple languages in Azure AI SearchCreate an index for multiple languages in Azure AI SearchMicrosoft · Documentation ·
Editorial approach
AI-assisted original writing, checked against linked public sources. Practical recommendations and illustrative scenarios are editorial analysis, not reported client results. Source dates and product details may differ from this collection’s preparation date.