Never Send a Human to do an Agent's Search
summary
Claim: agents succeed or fail on data, not models. Demonstrated build: an open source Lance file/table format that stores multimodal assets, embeddings, and metadata together and provides random access, secondary indexes, and time travel/versioning so searches are reproducible. Measured surprises: agentic queries fan out into hundreds of parallel requests and error rates compound across steps. Tooling tradeoffs shown: vector DBs blow up cost and fragmentation, while storing indexes and blobs in object storage cuts cost and simplifies iterative, high‑scale agentic search.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:45Agents will change everything
- 4:14What defines an AI agent
- 6:21Data as key agent differentiator
- 8:12Current data infrastructure limits
- 10:37Scale and multimodality challenges
- 16:23Iterative search and error accumulation
- 21:19Performance and evaluation importance
- 23:08Lance addresses agentic search limits
- 33:28Lance’s multimodal scalable features
- 36:21Agents as assistance
- 38:06Q&A
related talks
40mBuilding Intelligent Search Systems With Agentic RAG and OpenSearch
31mAgents Need More than Vector Search
31mAI Coding Agents Don't Need Bigger Models. They Need Better Tools.
28mAgentic AI
30mBuilding Robust, Secure LLM and Agentic AI Workflows
36mBuilding Next-gen Search With OpenSearch
41mAI is here – time to throw away our search engines?
12mEnterprise Agents Have a Structure Problem