Annotating enzyme function and finding candidate enzymes for a given reaction are essential steps for linking genomic data to biochemical networks. Existing approaches rely on sequence or structural similarity or on supervised training with known enzyme‑reaction pairs, which makes it hard to integrate biochemical evidence across large candidate spaces and to handle families with highly similar members.\n\nRIVER (Reaction‑enzyme Inference via Evidence‑guided Retrieval) is a multi‑agent framework that supports bidirectional retrieval: it predicts reactions that a given enzyme may catalyze and ranks enzymes that could perform a given reaction. The system orchestrates large language models, specialized biochemical computation tools, and external knowledge resources to gather evidence at several levels, including reaction chemistry, protein sequences and domains, catalytic mechanisms, and database annotations. Dedicated validation modules then examine evidence gaps and conflicts, and re‑rank candidates using traceable evidence records.\n\nOn public benchmarks, RIVER achieves the highest success rates for both enzyme retrieval and enzyme annotation. In family‑specific retrieval experiments, it outperforms all baselines on the selected datasets and metrics. For Escherichia coli enzyme retrieval tasks built from recent literature, RIVER delivers the best results with both closed‑source and open‑source model configurations, substantially surpassing closed‑source general‑purpose and biology‑specialized multi‑agent systems such as Claude Code, Codex, and Phylo Max.\n\nThese findings suggest that evidence‑guided multi‑agent retrieval can prioritize biochemically plausible enzyme‑reaction relationships, providing a solid basis for subsequent experimental validation.\n\nReview: RIVER’s integration of LLMs with domain‑specific tools showcases the power of evidence‑driven retrieval in enzyme function prediction and points to a promising direction for interdisciplinary computational biology.