Teaching a Reranker the Language of Security Tickets (+41% MRR@10) (opens in new tab)  🤖AI

TL;DR Our SOC's RAG pipeline retrieves over 142,000 closed XSOAR security tickets to ground investigation answers. After exhausting the easy wins — chunking, top-k, reranker choice — we still saw the right historical ticket land at rank 5-10 too often, and the LLM grounding its answer in a near-miss neighbor. We fine-tuned the reranker on our own data. Held-out test set, time-based split: MRR@10 BAAI/bge-reranker-v2-m3 (off-the-shelf) 0.598 Fine-tuned on 24K XSOAR pairs 0.846 +41% uplift. No ...