BioRAG-DRAG: A Multimodal Biological Retrieval Layer for Local-First Biomedical Agents

Biomedical agents need reliable access to heterogeneous evidence: literature text, gene and pathway records, protein sequences, DNA/cDNA sequences, and structured biological relations. Classical sequence tools such as BLAST remain the right choice for alignment-grounded verification, but they are not a unified context interface for large language model agents. We present BioRAG-DRAG , a local-first multimodal retrieval layer that combines pluggable neural sequence-text retrieval, BLAST verification, and graph-based evidence packaging. Specialized encoders such as ESM-2 can serve protein partitions, while OmniGene CPT provides a unified biological-language backbone for mixed sequence/text and agent-facing use; BLAST reranks or verifies sequence candidates; and DRAG graphs expose typed, traceable paths for downstream agents.

We introduce BioRAG-Standard v0 , a partitioned corpus/library with 257,886 retrievable records and an initial annotation layer for engineering evaluation built from Open-Rosalind Standard biomedical records and sequence-window extensions. On an in-index sequence-window stress test, BLAST nearly saturates biological matching, while vector retrieval recovers substantial but lower biological match rates. On held-out parent-fragment controls, public protein encoders outperform the current OmniGene protein-window embedding, while DNA/cDNA dense retrieval remains weak even with off-the-shelf Nucleotide Transformer pooling; this supports a model-agnostic BioRAG design rather than a claim that one unified generator backbone is the best sequence-search encoder. Indexed Chroma lookup over Standard text and 100k sequence-window collections adds only small lookup overhead after query embedding; this does not measure end-to-end instant latency. Finally, exploratory sequence DRAG traces show inspectable biological neighborhoods, including immunoglobulin-family and gene-symbol modules, with initial graph controls indicating non-random but partly sequence-similarity-driven structure. These results support a bounded architecture: vector retrieval supplies unified candidate context, while BLAST and DRAG provide biological verification and evidence attribution.