PRODUCTION SCENARIO
A home-improvement retailer's site search returns nothing for “stop my faucet dripping” even though it sells washer kits described as “repairs leaking taps”. Product text and shopper queries rarely share the same words.
Which capability closes this gap?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Text embeddings for search
Dense vector embeddings represent what a piece of text means rather than which words it contains, so retrieval can surface passages that align with the shopper's intent even when the wording differs completely - which is the whole problem when customers say faucet and the catalog says tap. Fine-tuning is the common first instinct: it changes what a model generates, while the site still needs a way to find the right product page. A larger context window governs how much a model reads in one request and does nothing about how candidates are retrieved. Batch inference is a real and useful way to process the whole catalog cheaply, and it is a scheduling and pricing choice rather than a matching technique - it would generate the embeddings, not replace them.