Navigating the Skies of Open AI: Perfect Your Grounding for a Smooth Landing
summary
Explains how to ground LLMs with retrieval using vector embeddings and concrete tooling: uses Azure AI Search and Blob Storage to index brochures, creating 3 PDFs into 12 chunks and returning matching chunks to the model; recommends chunking, overlapping, and summary embeddings to avoid the ~8k token input limit. Notes embedding sizes (e.g. text-embedding-ada-002 ≈1500 dims, OpenAI embedding-3 large ≈3000) and cost issues including a cited training cost ~$700M for large models, and shows a working demo web app.
AI-written summary · about
chapters (9)
- 0:00Open AI and grounding
- 3:19How LLMs work
- 8:37Limitations of base LLMs and their solutions
- 13:18Retrieval augmented generation
- 18:11Vector and hybrid search
- 23:55Chunking large documents
- 30:54Advanced retrieval and costs
- 36:58Private jet chatbot demo
- 48:21Q&A
related talks
48mUnderstanding Embedding Vectors and the RAG Pattern in AI Solutions
22mGenerative AI application stack and providing long term memory to LLMs
07mLightning Talk: Grounding LLM to Avoid Hallucinations Using Vector S...
01h 00mUnleashGPT: Mastering the Implementation of Large Language Models
28mUse OpenSearch as Your Vector Search Database To Optimize GenAI Accuracy!
43mBird's Eye View of Open-Source AI Infrastructure
48mUsing Azure AI Foundry to manage all your Large Language Models
16mEmpower AI: Building Smarter Solutions with Embeddings