Evidence-grounded retrieval and structured extraction for extraterrestrial organic matter literature

Extraterrestrial organic matter (EOM) records early Solar System chemistry and prebiotic processes and is central to biosignature interpretation. However, relevant evidence remains dispersed across PDF articles, tables, supplements, and scanned reports, limiting systematic reuse. We developed an evidence-grounded retrieval and structured-extraction workflow that converts 670 English-language PDFs into a queryable, traceable evidence layer. The corpus spans meteorites, returned samples from Ryugu and Bennu, and lunar materials, alongside literature on analytical methods and contamination control. The workflow combines optical character recognition (OCR)-compatible ingestion, Okapi Best Matching 25 (BM25) lexical retrieval, transformer-based dense retrieval, weighted hybrid fusion, optional cross-encoder reranking, evidence-grounded question answering, and schema-guided extraction. Evaluation on 45 queries and 683 judgments showed complementary retrieval behavior: hybrid retrieval provided the highest recall, BM25 the strongest reciprocal-rank performance, and dense retrieval the highest normalized ranking quality; no method dominated all metrics. All 670 PDFs were successfully ingested. Across 20 questions and five pipelines using the same large language model (LLM), the LLM-only condition provided no corpus-linked evidence, whereas all retrieval-augmented generation pipelines returned traceable source identifiers and 90%–95% evidence-checkable answers. Extraction was more reliable for meteorite names and organic compounds than for contamination controls, revealing remaining boundary and context challenges. These results establish a reproducible computational workflow for organizing traceable evidence from EOM literature, supporting biosignature interpretation and sample-return studies. Remaining limitations include OCR-heavy documents and table/figure content.