Reduce LLM calls with Vector Search Design Patterns
summary
Shows concrete, runnable patterns using Redis as a vector DB (RedisVL library, Redis OM Spring, RedisInsight) to cut LLM calls: semantic classification, semantic tool calling, and semantic caching. Provides demos (Kotlin notebooks, Testcontainers, sentence-transformer all-MiniLM-L6-v2 embeddings, LinkAsh embed for negation), loads 346 precomputed FAQs, and measures wins: cached response in ~92–345 ms vs full LLM runs of ~13 s and nonzero tokens. Cites model context/precision tradeoffs (256k→90%/80% then 1M tokens →36%) and reports ~5 billion tokens saved in January.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 4:03Redis beyond cache
- 12:16Vector search basics
- 12:37Semantic classification
- 22:12Semantic tool calling
- 30:20Semantic caching
- 40:50Spring AI integration
- 47:45Accuracy benchmarks
- 50:34Closing
related talks
46mLLM—Search Engine Killer?
40mTechniques to avoid common LLM pitfalls
50mKa-Ching! How Enterprises Measure LLM Performance and Cost
22mPractical Strategies for Implementing Domain-Specific LLMs
11mWhen Less Is More: A Technical Overview of LLMs and the Strength of Smaller Models
33mDesigning LLM Applications for Massive Datasets
55mVector Similarity Search in Spring with Redis Stack
22mIntegrating LLMs with Traditional Data Analysis