In materials research, composition, structure type, processing, and property evidence are often scattered across different repositories, so it is useful to test a retrieval task that connects them on a small scale first. For example, for a question such as “What process improved high-temperature stability in this composition family?”, the first retrieval can find candidate composition/structure families, and the next retrieval can be divided into two or more stages to find process records or literature evidence.
Rather than starting with large documents, it is better to begin with a narrow set of materials such as internal lab notebooks, paper abstracts, and descriptions of owned datasets. If you weakly label the correct answer at each retrieval stage with a person or a rule, it becomes easier to check whether the retrieval path is actually useful.
This approach is especially promising when materials knowledge is not complete in a single document and multiple pieces of evidence must be connected. However, it is safer to define the goal as “finding the next evidence to examine well” rather than “generating the answer.”