Conversational Search with LLMs
summary
Claim: LLMs let search move from single queries to long-term memory and multi-step, multi-channel flows. Concrete tools and patterns noted: a multi-tiered query understanding stack (cheap NLP for simple queries, LLMs for hard ones), semantic caching, fine-tuned embeddings and multi-stage retrieval+rerank, and query-focused summarization. Built features include task decomposition for loans and real-estate workflows; verification by decomposing outputs into atomic chunks and checking against databases. Measured impact: clarification-driven flows produce up to ~20% conversion lift, latency and cost remain practical constraints.
AI-written summary · about
chapters (9)
- 0:00Conversational search intro
- 1:06Differences in conversational search
- 3:58Search clarification importance
- 8:58Query understanding beyond semantics
- 13:55Embedding models for vertical search
- 17:45Ranking models and relevance
- 20:59Summarization challenges
- 22:42Hallucination risks in LLMs
- 23:46Future directions
related talks
49mLLM—Search Engine Killer?
51mQuery Understanding With Large Language Models: Techniques and...
21mTransforming search and discovery using LLMs
01h 30mUsing LLMs to improve your Search Engine
46mLLM—Search Engine Killer?
20mTransforming the Search Industry with LLM Using Source Frontier
22mPractical Strategies for Implementing Domain-Specific LLMs
36mOptimizing LLM Performance With Caching Strategies in OpenSearch