In this article, Wang et al. developed and tested a large language model (LLM) framework for extracting suicide-related Social Determinants of Health (SDoH) from unstructured narratives such as clinical notes. The framework involves two steps, context retrieval and relevance verification, after which SDoH factors are extracted. During context retrieval, input text is split into individual sentences, and a pre-trained LLM identifies which sentences contain the SDoH factor of interest. In the relevance verification step, chosen sentences are given to another LLM which verifies their relevance. These sentences are then given to an LLM that identifies and extracts specific SDoH factors occurring within two weeks before suicide. This framework was compared to a fine-tuned BioBERT model, a GPT-3.5-turbo End-to-End model, a GPT-3.5-turbo Chain-of-Thought model, and DeepSeek-R1, and data from the 2020 version of the National Violent Death Reporting System was used. A pilot user study was also performed where participants were tasked to annotate selected SDoH factors across incidents with and without the relevant context highlighted by the framework. The framework outperformed all baseline methods in 9 out of 10 infrequent suicide-related SDoH factors and improved F-1 scores for 5 out of 8 frequent SDoH factors. DeepSeek-R1 outperformed other models in 7 out of 16 SDoH factors, but the proposed framework outperformed it on recall for 11 factors. The pilot user study showed that the AI-assisted method was faster and slightly more accurate and that it was preferred among participants. The results suggest that the proposed framework may be suited to high-coverage extraction tasks while reasoning models like DeepSeek-R1 are preferable for minimizing false positives.