Techniques to avoid common LLM pitfalls
summary
Demonstrates concrete fixes: use Gemini 1.5: 2 million token context for huge inputs, ground outputs with Google Search or Vertex AI Search, or run a RAG pipeline with embeddings and a vector DB (example uses LangChain + Firestore vector store) to answer from private PDFs. Showcases function calling to hit real APIs (weather example), response schemas to force JSON, context caching and batch generation 50% cheaper for cost savings, and evaluation with DeepEval and Promu; 10 tests, 8 failures to catch regressions and safety holes.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:28Common LLM pitfalls
- 8:13Problems with LLMs and their limitations
- 12:05Avoiding LLM pitfalls
- 13:28Grounding LLMs
- 16:17Retrieval Augmented Generation
- 24:08Function calling for data
- 27:58Controlling output format
- 29:44Reducing costs with caching
- 32:33Evaluating LLM performance
- 36:03Testing LLM outputs
- 39:16Summary
related talks
36mTechniques to avoid common LLM pitfalls
45mAvoid common LLM pitfalls
51mAvoid common LLM pitfalls
57mLLMs gone wild
01h 49mAI engineering with open access LLMs that lie, curse, and steal
48mGoing Beyond Prompt Engineering
49mRAG: from dumb implementation to serious results
28mAPIs and GenAI : Your Perfect Combo with LLM-Powered Frameworks.