Building a Semantic Search Engine
summary
Shows a working pipeline: index 25,000 images with captions, ingest at ~3,000 docs/sec into BM25 baseline ElasticSearch with ~4 ms retrieval, compute sentence embeddings via sentence-transformers (takes ~3–4 minutes on CPU for the corpus), store vectors in a vector DB and run FAISS/ANN searches that cut recall/latency tradeoffs (flat ≈10s ms, ANN ≈10 ms on CPU faster on GPU). Embedding retrieval improves relevance over BM25 in examples.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:37Search systems evolution
- 10:13Token-based retrieval with Elasticsearch
- 23:46Limitations of token-based
- 24:48Embedding-based retrieval intro
- 30:08Training models for similarity
- 45:05Training embeddings and negatives
- 50:08Multimodal search with CLIP
- 58:57Vector indexing strategies
- 1:09:23Building vector indexes
- 1:21:23Multi-stage retrieval systems
- 1:26:05Production trade-offs
related talks
01h 30mBuilding a Search Engine
39mAI-powered Semantic Search; A story of broken promises?
22mThe Capabilities of Semantic Search
32mBuilding semantic search using Deep Learning
13mA Beginner’s Guide to Semantic Search
32mRetrieval Augmentation and Semantic Search at Scale
46mNatural Language & Search, Improve your Search w/ Vector capabilities
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline