Neural Search - Let's talk about quality
summary
An open-source finetuner is used with Elasticsearch plus vector search to improve e‑commerce fashion retrieval on a 410‑item WDC corpus; plain CLIP performs poorly then jumps (recall roughly 10→40) after full fine‑tuning, while a text encoder (“ms market stillboard base v3”) moves only slightly. Training uses permutations of attributes against product title/description with triplet margin loss and separate query/document encoders, and final ranking combines bm25 + expert + 2x clip scores. The pipeline is built, reproducible, and yields clear recall gains over BM25 alone.
AI-written summary · about
chapters (13)
- 0:00Introduction and calibration
- 1:24What is neural search
- 3:53Why fine-tuning is necessary
- 6:00Experiment setup and data
- 9:46Model performance results
- 14:43How we trained the models
- 17:55Neural vs vector search
- 20:44Quality focus in neural search
- 21:47Summary and open source
- 22:09Hiring for implementation
- 23:31Fine-tuning depth discussion
- 24:57Combining neural and standard search
- 27:24Q&A session
related talks
46mNatural Language & Search, Improve your Search w/ Vector capabilities
22mMachine Learning: The Engine Behind Modern E-commerce Search
32mBudget Friendly Semantic Search With Neural Sparse Search
39mModel Fine-tuning For Search: From Algorithms to Infra
33mAI-Powered Search in Modern E-Commerce
31mIntelligent Japanese Search With OpenSearch
01h 09mVector Search: Ask Me Anything!
39mAI-Powered Search in Modern E-Commerce