archive – Andrei Lopatenko
summary
Twenty years of search experience frames the claim that conversational search needs long-term memory and multi-step, multi-channel flows. Empirical specifics: clarification appears in roughly 30% of sessions, drives up to 20% conversion uplift, and users tolerate extra latency beyond the usual several-hundred-millisecond SLAs when saving effort. Engineering realities: one-million-QPS scale makes LLM-only inference impractical, so use multi-tiered NLP, semantic caching, fine-tuned embeddings, RAG plus fact-checking by decomposing outputs and verifying against indexes.
AI-written summary · about
chapters (10)
- 0:00Conversational search intro
- 1:08Conversational search differences
- 3:13Search trends and user expectations
- 5:02Search clarification role
- 9:32Query understanding beyond embeddings
- 14:04Fine-tuning embeddings for vertical search
- 16:33Multi-stage pipelines for relevance
- 18:39Summarizing search results
- 21:51Mitigating LLM hallucinations
- 22:55Future of conversational search
related talks
30mBeyond Keywords: Building an AI-Powered Search That Reads Your Mind
51mQuery Understanding With Large Language Models: Techniques and...
38mSearch is Back: Solving the "Context Crisis" for AI Agents
20mTransforming the Search Industry with LLM Using Source Frontier
22mAgentic Hybrid Retrieval: Orchestrating LLM Agents, Hybrid Search & Explainability in OpenSearch
01h 30mOptimizing Retrieval for Agentic Systems
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline
32mOpenSearch Agentic Memory: Intelligent Context for AI Agents