RPTune: Learned Context Curation for LLM Catalog Search
Chuxuan Hu, Hejie Cui, Norman Huang, Shubham Kumar Bharti, Wang-Chiew Tan, Sercan Ö. Arık
Google · University of Illinois Urbana-Champaign
Published October 1, 2026 · on the feed October 2, 2026 · ▲ 32 · arXiv 2610.00964
The abstract
For small merchant businesses (SMBs) whose catalogs fit within a long-context LLM, full-catalog prompting offers a compelling alternative to multi-stage retrieval designed primarily for large marketplaces with millions of items. However, fitting the full catalog into the context window does not ensure that the model can use it effectively, since LLMs do not exploit long contexts uniformly. We therefore study in-context catalog search through two complementary questions: (1) how to curate and present catalogs to the LLM, and (2) how to adapt the LLM for product selection on curated contexts. We propose RPTune, an end-to-end framework that couples learned catalog curation with LLM post-training using automatically generated, catalog-grounded supervision. An encoder-reorganizer curator orders and prunes products guided by downstream LLM feedback, while the resulting curated catalogs in turn improve the effectiveness of LLM post-training with a context-relative reward. We evaluate RPTune on 7 real merchants spanning distinct retail verticals, using 100 complex conversational queries per merchant. RPTune consistently improves search accuracy across both proprietary and open-weight LLMs, with context curation yielding gains of up to 31.4 percentage points and post-training adding a further 10.3 points on average.
The Brief
2:05A short podcast summary of the paper. Written and voiced by AI.
Transcript
RPTune. Learned Context Curation for LLM Catalog Search. By Chuxuan Hu and five others, from Google and the University of Illinois Urbana-Champaign. Published on the first of October. No code repository is linked to it.
Picture a small merchant business whose whole catalog fits inside a long-context language model. For a business like that, prompting with the full catalog is a compelling alternative to multi-stage retrieval, which was designed mainly for large marketplaces with millions of items. But fitting the catalog into the context window does not mean the model can use it well, because language models do not exploit long contexts uniformly. Some parts of a long context get used better than others. That leaves two complementary questions. How should a catalog be curated and presented to the model? And how should the model be adapted to select products from a curated context?
RPTune answers both in one end-to-end framework. It couples learned catalog curation with post-training of the language model. The supervision that drives it is generated automatically, and it is grounded in the catalog itself, so the training signal comes from the very products the model will be searching.
The curator has an encoder-reorganizer design. It orders and prunes products, deciding what goes in and in what sequence, guided by feedback from the downstream language model. The curated catalogs then make post-training more effective, with a reward defined relative to the context. Each half improves the other: the model's feedback shapes the curation, and the curation shapes what the model is trained on.
Evaluation covers seven real merchants spanning distinct retail verticals, with one hundred complex conversational queries per merchant. RPTune consistently improves search accuracy, on proprietary and open-weight language models alike. Context curation yields gains of up to thirty-one point four percentage points, and post-training adds a further ten point three points on average on top of that.
For a catalog that already fits in the window, the question shifts from what to retrieve to how to arrange it. The arrangement is learned from the model that has to read it, and the model is then tuned to read that arrangement well.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here the whole catalog fits, yet the model still misses the right product.
You run product search for a small online shop. Its whole catalog fits inside a long-context language model, so you paste every product into the prompt along with the shopper's question.
The road ahead
- 1 · The problem ~6 min A small shop's whole catalog fits in a language model's prompt, but the model does not read every part of a long prompt equally well.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min A learned curator trims and orders the catalog for each query, trained on the model's own picks, and the model is then post-trained on those curated contexts.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Make training data from the catalog, embed query and products, train a reorganizer on model feedback to prune and order, then post-train the model.
- 6 · The experiment ~6 min You’ll guess: At most, how many percentage points of search accuracy did context curation alone add?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.