Transforming search and discovery using LLMs
summary
Instacart replaces brittle engagement-trained components with LLM-backed modules for query classification and rewrites, precomputing head/torso outputs and batching generations to avoid latency. A hybrid flow feeds top-converting categories and query metadata into LLMs, yielding much higher tail performance — 18 percentage points precision and 70 percentage points recall on tail queries — plus fewer zero-result searches, generated substitute and complementary carousels, and plans to deploy a distilled Llama model for real-time tail coverage.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:03Importance of search in grocery
- 2:06Challenges in conventional search
- 4:12Using LLMs for query understanding
- 8:52Query rewrites with LLMs
- 10:34Serving LLMs efficiently
- 13:22LLMs for product discovery
- 18:44Key takeaways and Q&A
related talks
20mBeyond Relevance: Using LLMs to Transform E-commerce Search into Product Discovery Engines
01h 30mUsing LLMs to improve your Search Engine
24mConversational Search with LLMs
46mLLM—Search Engine Killer?
49mLLM—Search Engine Killer?
20mTransforming the Search Industry with LLM Using Source Frontier
25mTransforming E-Commerce with LLMs: 3 Real-World Use Cases
51mQuery Understanding With Large Language Models: Techniques and...