The Art of Production Quality Retrieval
summary
Walks through a production retrieval pipeline: generate captions, mood keywords and image/text embeddings, parse queries with an semantic router LLM to route intent, then run hybrid search over multiple embedding columns plus filters and apply reranking. Claims: baseline hit@5 ≈ 48.75% and six epochs of an embedding fine tune raises hit@5 by ~10%; reranking can add up to ~12% at top five. Practical notes: rerankers need a few hundred labeled pairs, embedding fine-tunes need thousands (10k+ yields solid lift), and synthetic data often fails.
AI-written summary · about
chapters (16)
- 0:00Production quality retrieval intro
- 2:52Four key retrieval questions
- 4:16Demo: semantic.art search
- 7:52Balancing user intent and features
- 9:13Generating searchable artwork features
- 11:20Embedding search space construction
- 13:10Fine-tuning embedding models
- 14:53Semantic routing for user intent
- 16:21Hybrid search query pipeline
- 17:24Reranking methods for results
- 20:19Fine-tuning for hit rates
- 21:42Iterative improvement loops
- 22:24Demo to production summary
- 23:05Q&A: Data quantity vs. quality
- 25:43Q&A: Dataset creation pitfalls
- 27:10Q&A: Ranking metrics
related talks
13mThe State of AI Powered Search and Retrieval
01h 30mBuilding a Search Engine
39mAI-Powered Search in Modern E-Commerce
38mEmbeddings! Embeddings everywhere!
01h 28mBuilding a Semantic Search Engine
40mExploring reranking depth in modern search pipelines
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
19mHow to look at your data