RPTune: Learned Context Curation for LLM Catalog Search
Long context alone did not make catalog search work; learned curation did.
RPTune targets SMB catalogs small enough to fit inside an LLM prompt, then learns how to order and prune the products so the model can use that context better. The framework pairs an encoder-reorganizer curator with LLM post-training built from automatically generated, catalog-grounded supervision. Across 7 real merchants and 100 complex conversational queries per merchant, the paper reports accuracy gains across proprietary and open-weight models. Context curation improved accuracy by up to 31.4 percentage points, with post-training adding another 10.3 points on average. HF Daily Papers' note
RPTune targets SMB catalogs small enough to fit inside an LLM prompt, then learns how to order and prune the products so the model can use that context better. The framework pairs an encoder-reorganizer curator with LLM post-training built from automatically generated, catalog-grounded supervision. Across 7 real merchants and 100 complex conversational queries per merchant, the paper reports accuracy gains across proprietary and open-weight models. Context curation improved accuracy by up to 31.4 percentage points, with post-training adding another 10.3 points on average. HF Daily Papers' note
score 4