Chemistry retrieval model Valence targets near-miss passages: 81.8% hit rate in-field, versus 2% to 4.5% for general models

AI retrieval for chemical R&D hinges not on cross-field sorting but on separating passages within one field that share vocabulary, material categories and even patent classifications. General-purpose retrieval models are built for breadth and can tell a coatings document from a personal care document, yet often cannot judge which of two same-field passages answers the question. In one real-corpus query, a leading general-purpose retrieval model ranked the correct passage beyond position 300 out of roughly one million and put the wrong one first, giving the two nearly identical scores.

Valence was trained on tens of thousands of near-miss pairs, learning the signals a formulation chemist uses to tell similar passages apart, such as an ingredient's role and the application it targets. General models squeeze a field's passages into a narrow score band, leaving little difference between first and hundredth place; Valence spreads them apart so the right answer rises to the top. FormBench built a corpus of roughly one million passages and more than 55,000 chemist-style queries from over 590,000 US formulation patents, each query paired with the passage holding the answer; the hard test draws distractors from the same field.

The reported evaluation is FormBench: roughly one million passages and more than 55,000 queries built from over 590,000 US formulation patents, spanning pharmaceuticals, adhesives, coatings, agrochemicals, food and personal care. Valence is described as the retrieval engine behind Ask Albert and can be tuned on a company's own documents.

The report says some general-purpose models dropped by more than 80% as the corpus grew toward one million passages while Valence held steady, but gives no specific Valence figures across corpus sizes. The hard-test hit rate and strict metric are as reported and have not been independently reproduced.