Avoid common LLM pitfalls
summary
Covers concrete fixes: use frameworks like LangChain with PDF splitting into 500‑char chunks and 100‑char overlap, embed into a vector DB (example uses Firestore) and run Retrieval-Augmented Generation (RAG) so a PDF yields the correct cargo capacity (13.5 cu ft) instead of hallucination; enable grounding via Google Search or Vertex AI Search; grant models realtime data with Function Calling; exploit Gemini 1.5 Pro’s 2 million token context window and batch generation for ~50% cost savings.
AI-written summary · about
chapters (14)
- 0:00Introduction to talk
- 1:26LLM landscape overview
- 6:20Common LLM pitfalls
- 13:05Solutions for pitfalls
- 15:27Pre and post-processing fixes
- 16:42Grounding LLM responses
- 21:25Retrieval Augmented Generation
- 27:00Function calling for realtime data
- 31:18Controlling LLM output format
- 34:45Optimizing LLM costs
- 38:35Observability for LLM calls
- 41:44Evaluating LLM outputs
- 45:19Securing LLM applications
- 49:53Talk summary and Q&A
related talks
36mTechniques to avoid common LLM pitfalls
40mTechniques to avoid common LLM pitfalls
57mLLMs gone wild
45mAvoid common LLM pitfalls
18mLLMs and GenAI: A Guide to Building Your Own RAG Model for Document Interaction
01h 33mRAG How To Talk To Your Data
29mUsing LangChain and LlamaIndex to customize LLMs...
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance