RAG vs Fine tuning which one is better for your AI project
summary
Argues RAG wins for most apps because it retrieves fresh facts via embeddings and a vector DB (Pinecone, pgvector) while fine-tuning only buys consistent, structured outputs and needs hundreds to thousands of labeled examples and privacy scrubbing. Describes a working pipeline: Mail Magic pipeline embeds sent emails, semantic-searches matches, extracts few-shot examples and yields better replies; fine-tuning is used only when exact personal wording is required. Notes structured SQL queries can replace vectors for highly tabular data and cites chess filters (≈2,000 ELO threshold) as a concrete example.
AI-written summary · about
chapters (14)
- 0:00Introduction and background
- 3:53Fine-tuning overview and use cases
- 7:25Retrieval Augmented Generation explained
- 13:47RAG vs fine-tuning for email replies
- 16:42Retrieval with structured data
- 19:34Chess puzzle example with SQL queries
- 24:56Fine-tuning for query consistency
- 27:26Prompt engineering vs fine-tuning
- 31:21Prompt validation and optimization tools
- 33:54Adapting email tone for new roles
- 36:20RAG vs prompt context for startups
- 39:32Testing strategies for AI applications
- 41:14Prompt validation tools overview
- 43:00Q&A and wrap-up
related talks
19mRAG in 2025: State of the Art and the Road Forward
30mRAG: Tricks from the trenches.
09mFine-tuning embeddings 4 vector search
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
22mOptimizing AI: From RAG to Riches!
21mHow RAG is Redefining AI Capabilities
53mBuild RAG from Scratch
49mRAG: from dumb implementation to serious results