The Three-Body Problem of Inverse Hybrid Search
summary
Solves saved-image alerts by turning product arrivals into queries against millions of stored searches, combining vector+boolean+lexical matching with strict fetch-all semantics so zero matches are missed. Uses Elastic percolator concepts augmented with Vespa predicate fields and ranking to support image embeddings against a real client corpus of 170 million listings. Finds correctness needs a golden evaluation set, cardinality drives P95 match-set sizing, and execution physics (scatter/gather, payload bytes, queuing) dominate costs; the prototype works in production and stabilizes when all three axes are measured.
AI-written summary · about
chapters (11)
- 0:00Introduction and background
- 2:59Saved search use case
- 6:40Inverse search concept
- 10:38Forward vs inverse search
- 13:13Three-body problem overview
- 14:41Gem app use case
- 17:16Correctness challenge
- 20:09Cardinality challenge
- 25:57Execution physics challenge
- 31:28Observability and resilience
- 33:20Q&A session
related talks
37mEnhancing AI and LLM Systems with Hybrid Search
22mAgentic Hybrid Retrieval: Orchestrating LLM Agents, Hybrid Search & Explainability in OpenSearch
51mQuery Understanding With Large Language Models: Techniques and...
47mBoosting AI With OpenSearch: Vector Stores, Similarity Search, and Retrieval-Augme...
49mHybrid search on hybrid models, at scale
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
01h 30mBuilding a Search Engine
30mBeyond Keywords: Building an AI-Powered Search That Reads Your Mind