What is Retrieve-for-Train?

Retrieve-for-Train is a Google Research method for AI search that does the hard planning once, during training. At search time a small model then turns one query into a full, varied set of results in a single pass. The problem it solves is the search that needs a set, not a single best match. Someone who types "camping gear" wants a tent, a sleeping bag, a stove and a lamp, not ten tents. Working that set out live, with a large AI model reasoning step by step while the user waits, is slow. Retrieve-for-Train moves that reasoning into a one-off training run and hands the result to a much smaller model, which Google reports answers 12 to 20 times faster.

Retrieve-for-Train is Google's way of getting a fast, varied set of search results without making an AI model think at search time. Google Research published it in September 2026. The problem starts with a broad query. Someone searches for "camping gear", and a good answer is a set that covers the trip: a tent, a sleeping bag, a stove, a headlamp. Search systems get that set by splitting the one query into several smaller ones and running each. The obvious way to do the splitting is to hand the query to a large language model at search time. That fails in two ways. The model tends to produce near-copies of one query instead of queries for different kinds of item. And it works out its answer step by step, in text, so it is slow: in Google's tests, splitting 1,024 queries at once took nearly 50 seconds. Retrieve-for-Train does the thinking once, in three steps. First, a medium-sized language model is trained by trial and reward to produce 10 smaller queries per search. The reward scores the 10 as a group: do they match real items in the catalogue, do they cover different ground, and do they stay on the original topic. Second, that trained model writes out huge numbers of examples, each pairing a query with a good set of results, with no human labelling. Third, a small model of about 54 million parameters learns from those examples to jump from a query straight to the whole set of results at once, rather than word by word. Only the small model runs when a user searches. In Google's tests it took under a second for small batches and about 4 seconds for the batch of 1,024, 12 to 20 times faster than a step-by-step model. Its results were also more varied than a general model's. It is a recipe for compressing one expensive behaviour into one cheap model, not a new search engine.

A gift-hamper business has a problem. A customer who orders a "picnic hamper" wants a spread, not six kinds of cheese, and every hamper has to be varied, on theme and made of things actually in the storeroom. One way to run it is to hire a clever generalist who invents each hamper from scratch while the customer waits. That is slow, and the generalist keeps reaching for five near-identical crackers. The owner instead spends 4 weeks working out, for each kind of order, which items make a good hamper, and scores each attempt on what matters: real stock, variety, and fit with the order. The results go onto assembly cards. From then on, any new hire can fill a hamper in 60 seconds without thinking, and the hampers are better than the generalist's. Retrieve-for-Train runs a search system the same way. The 4 weeks of working out are the training, the assembly cards are the examples, the new hire stands in for the small model, and the hamper for the set of results. Where the picture breaks: a new hire follows a fixed card for a fixed order, while the small model has learned the pattern and produces a set for orders it has never seen.