Even in the era of large language models (LLMs), biomedical relation extraction (RE) still plays a major role in timely creation of knowledge graphs that further guide biomedical knowledge discovery. The main task in RE is to extract a relation "as expressed" in an input text. At times, crucial definitional information or other auxiliary information about the entities involved may be missing from the input text. Augmenting it from other external textual sources appears helpful on the surface but can be harmful too, as these sources can overwhelm the signal in the original input, leading to false positives or false negatives. To counter this, we leverage a pre-trained biomedical text retriever to augment original inputs with additional instance-specific snippets. This is done through a gating mechanism that allows the retrieved snippets to enhance but not overwhelm the signal from the original input. We evaluate our approach on three standard biomedical relation extraction datasets (CDR, BioRED, and ChemProt) and show consistent improvements (up to 10 F1 points) compared with strong supervised baselines involving both encoder and decoder models. All our code and the datasets used are available for reuse: \url{https://github.com/bionlproc/GRAFT-RE}.

Extracting relations from scientific literature is a fundamental task in biomedical NLP because entities and relations among them drive hypothesis generation and knowledge discovery. As literature grows rapidly, relation extraction (RE) is indispensable to curate knowledge graphs to be used as computable structured and symbolic representations. With the rise of LLMs, it is pertinent to examine if it is better to skip tailoring supervised RE methods, save annotation burden, and just use zero shot RE (ZSRE) via LLM API calls. In this paper, we propose a benchmark with seven biomedical RE datasets with interesting characteristics and evaluate three Open AI models (GPT-4, o1, and GPT-OSS-120B) for end-to-end ZSRE. We show that LLM-based ZSRE is inching closer to supervised methods in performances on some datasets but still struggles on complex inputs expressing multiple relations with different predicates. Our error analysis reveals scope for improvements.