60-80% of individuals at clinical high risk (CHR) for psychosis report lifetime suicidal ideation (SI) yet only 18% attempt suicide. Therefore, it is critical to differentiate individuals with suicidal behavior (SB), a strong predictor of a future attempt, from SI alone. Markers such as fearlessness, risk-taking, and abnormal threat and emotion responses differentiate SB from SI. Using a GoEmotions-trained BERT model, we previously found that CHR individuals with SI use speech more semantically similar to anger, an effect specific to those with SB after stratification. We now use Retrieval Augmented Generation (RAG) which grounds a generative model on open ended interviews from 81 CHR individuals (25 with SB, 6 with SI only, and 50 with neither) to map linguistic patterns onto clinically validated groupings from the Columbia-Suicide Severity Rating Scale. We examined if CHR speech with SB, relative to SI only, reflects lower threat response (Q1), risk aversion (Q2), fear (Q3), and emotion dysregulation (Q4). Of 4 queries, Q1 was supported. RAG identified that SB language more commonly contained detached descriptions of painful events and identified language reflecting downplaying of dangers and reduced concern for consequences, linguistic patterns that may reflect a reduced response to threat. Clinically, examining RAG-derived linguistic markers offers a scalable approach to identify CHR individuals at elevated suicide risk and may guide targeted prevention strategies.