RAG

Retrieval-Augmented Generation (RAG): LLMs retrieve relevant chunks from uploaded documents at each query.The Kaparthy LLM Wiki

Characteristics

• Knowledge rediscovered from scratch every question.
• No accumulation or persistence.
• Examples: NotebookLM, ChatGPT file uploads, most RAG systems.

Contrast with LLM Wiki

• RAG: Ephemeral retrieval.
• LLM Wiki: Proactive maintenance of persistent wiki.

From Retrieval-Augmented Generation

• RAG optimizes LLM output by referencing authoritative knowledge bases outside training data without retraining.
• Solves challenges like hallucinations, static knowledge cut-offs, and lack of source attribution.
• Pipelines involve creating external data (embeddings), retrieving relevant information (vector search), and augmenting the LLM prompt.
• Cost-effective alternative to fine-tuning for injecting domain-specific or real-time data.