Keynote: Building and Productionizing RAG
summary
Concrete recipe: a Python library called llama index ingests documents by chunking and embedding into a vector DB, then runs retrieval plus LM synthesis — often in five lines of code. Default knobs matter: chunk sizes around 1k tokens and top_k=2 are common starting points. Production adds small-to-big retrieval or a fusion retriever, agent wrappers, and fine-tuning of embeddings or LMs using synthetic queries. The ecosystem includes 150+ data loaders and integrations with 20+ LLMs.
AI-written summary · about
chapters (9)
- 0:00Building RAG systems
- 5:48How RAG works
- 13:16RAG challenges and failures
- 17:51Evaluating RAG performance
- 24:30Optimizing RAG techniques
- 36:29Advanced RAG retrieval
- 43:11RAG with agents
- 53:02Fine-tuning for RAG
- 57:42Productionizing RAG
related talks
42mTake Your RAG To The Max!
18mLLMs and GenAI: A Guide to Building Your Own RAG Model for Document Interaction
01h 12mBuild RAG using Llama 3.1
28mNavigating Challenges and Enhancing Performance of LLM based Applications
10mKeynote: Unlocking Corporate Data with Retrieval-Augmented Generation
35mAdvanced RAG Techniques with LlamaIndex
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
32mBuilding Multi-Agent RAG Systems