As the first stage of a materials literature search pipeline, it seems suitable to try a method that separately indexes title, abstract, methods, table, and figure caption, then narrows candidates with Boolean conditions. For example, one could prioritize only literature that includes metadata such as composition, temperature, pressure, synthesis method, and measurement method, and configure the agent to fetch only the sections it needs.
This approach may be especially well suited to tasks where “what evidence exists” matters more than “what is true,” such as drafting review articles, compiling comparison tables of experimental conditions, or collecting evidence for specific process conditions. However, because materials papers vary widely in notation, field extraction quality is likely to determine performance, so it would be better to compare against hand-labeled ground truth on a small corpus first.