Yuval Belfer

Every RAG system bets everything on a single chunk size. 500 tokens? 800? Pick wrong, and half your queries fail before they start. But here's what nobody tells you: all the picks are wrong; there is no single chunk size that works for all queries. We ran oracle experiments across meeting transcripts, story chapters, and TV scripts. The result? Queries disagree violently on what chunk size works best - sometimes by 40 percentage points. Your "tuned" chunk size isn't a compromise; it's systematic underperformance. In this talk, we'll expose why fixed chunking fails and show you a dead-simple fix: index at multiple chunk sizes, aggregate at retrieval time using Reciprocal Rank Fusion. No retraining. No LLM overhead. Just 1-37% better recall across benchmarks by letting queries vote with their ranks instead of forcing them into one-size-fits-all boxes. Walk away knowing exactly when your chunk size is sabotaging you - and how to stop leaving 20-40% of your retrieval performance on the table. Speakers: Yuval Belfer — AI21 Labs; Niv Granot.

Your model generates gibberish. Once every thousand prompts. High confidence scores. No crashes. No warnings. We hit this twice while building Jamba models. First: A request gets misclassified during scheduling, loads stale state from a previous prompt cache slot, and confidently generates nonsense. Second: Logprob spikes during RL training that looked like training instability-until we noticed they tracked with rollout count, then with cache size. In this talk, we'll walk through both debugging journeys-the false starts, how we instrumented vLLM to thread request IDs through the forward pass, the search for variables that change failure structure rather than magnitude, and the lesson both share: distributed inference systems fail silently. No stack trace. No sanitizer warning. Just wrong answers with perfect confidence. You'll learn how to build comparison scripts that expose logprob divergence, force memory pressure to surface rare bugs, and shrink a distributed RL training mystery into a reproducible single-script failure. Walk away knowing how to debug vLLM when it lies to you quietly. Speakers: Asaf Gardin — AI21; Yuval Belfer — AI21 Labs.