A stock-photo customer types a description and expects relevant images, not a tour through millions of categories. Similarity search addresses that problem by finding items that match a query and ordering them by relevance. The business question comes first: will a service handle text-to-image search, or does this particular marketplace need its own solution? section 3.4.2
Relevance and speed pull in different directions. A simple rule, such as matching query words, can identify plausible candidates quickly; a more sophisticated model can then rank a much smaller set. Language or location filters may narrow the pool earlier still. This two-stage design makes expensive ranking practical without applying it to every item. section 3.3
relevancy = f(encode_text(query), encode_image(item))
For text-to-image search, the scoring sketch compares an encoded query with an encoded image. It identifies what must be ranked without prescribing a particular encoder or scoring function. section 3.4.2
Check the retrieval stage against its actual job:
- Keep enough relevant candidates for ranking. In a stock-photo experiment, 97% of clicked photos appeared in the new model’s top 1,000, suggesting room to add a candidate-retrieval stage—not proving that any particular stage will succeed. section 12.4.2
- Watch for misleading matches in evaluation. A language-model validation procedure retrieves nearest training chunks for each validation chunk and measures token overlap, so near-duplicates can be filtered rather than rewarded as useful retrieval. section 7.3.3
A better match score is not the whole outcome. For photo search, also watch search time, result relevance, and whether sessions end in purchases. Those measures connect retrieval decisions to the customer experience and marketplace results. section 12.4.2
