rag

name	rag
description	RAG pipelines including document chunking, embedding, vector search, retrieval strategies, and reranking
layer	domain
category	ai-ml
triggers	["RAG","retrieval augmented generation","vector search","embeddings","chunking","knowledge base","semantic search","reranking"]
inputs	[{"documents":"Source documents to index (PDFs, web pages, databases, etc.)"},{"query_patterns":"How users will search (natural language, keyword, hybrid)"},{"requirements":"Accuracy, latency, freshness, cost constraints"},{"infrastructure":"Vector database choice, embedding model preference"}]
outputs	[{"pipeline_design":"End-to-end RAG pipeline architecture"},{"chunking_strategy":"Document splitting approach with overlap"},{"embedding_config":"Model selection and dimension settings"},{"retrieval_strategy":"Search, filter, and reranking configuration"},{"evaluation_plan":"Metrics and test cases for RAG quality"}]
linksTo	["ai-agents","prompt-engineering","postgresql","redis"]
linkedFrom	["ai-agents","cook","research"]
preferredNextSkills	["ai-agents","prompt-engineering"]
fallbackSkills	["research"]
riskLevel	low
memoryReadPolicy	selective
memoryWritePolicy	none
sideEffects	["May create vector store indexes","May process and embed documents","Consumes embedding API tokens"]

RAG (Retrieval-Augmented Generation) Skill

Purpose

Design and implement RAG pipelines that ground LLM responses in relevant, up-to-date knowledge. This skill covers the full pipeline: document ingestion, chunking strategies, embedding model selection, vector storage, retrieval with filtering, reranking for precision, and prompt construction with retrieved context. RAG reduces hallucination by giving the model factual context to work with.

Key Concepts

RAG Pipeline Overview

INGESTION (offline):
  Documents -> Chunking -> Embedding -> Vector Store

RETRIEVAL (at query time):
  Query -> Embed Query -> Vector Search -> Rerank -> Top-K Context

GENERATION (at query time):
  System Prompt + Retrieved Context + User Query -> LLM -> Response

EVALUATION:
  Context Relevance | Answer Faithfulness | Answer Relevance

Chunking Strategies

FIXED SIZE:
  Split at N characters/tokens with M overlap
  Simple but may break mid-sentence
  Good for: Uniform, well-structured documents

RECURSIVE CHARACTER:
  Split on paragraphs, then sentences, then words
  Respects document structure
  Good for: General purpose, articles, documentation

SEMANTIC:
  Split when embedding similarity between sentences drops
  Each chunk is semantically coherent
  Good for: Long documents with topic changes

DOCUMENT-AWARE:
  Split on markdown headers, HTML tags, code blocks
  Preserves structural meaning
  Good for: Technical documentation, code, structured content

RULES OF THUMB:
  Chunk size: 500-1000 tokens (sweet spot for most models)
  Overlap: 50-200 tokens (prevents losing context at boundaries)
  Smaller chunks = more precise retrieval, less context per chunk
  Larger chunks = more context per chunk, less precise retrieval

Embedding Model Selection

MODEL                    DIMENSIONS  SPEED    QUALITY  COST
text-embedding-3-small   1536        Fast     Good     Low
text-embedding-3-large   3072        Medium   Best     Medium
voyage-3                 1024        Fast     Great    Medium
cohere-embed-v3          1024        Fast     Great    Medium
BGE-large (open source)  1024        Self-host Good    Free (compute)

GUIDELINES:
  - Start with text-embedding-3-small for prototyping
  - Upgrade to text-embedding-3-large or voyage-3 for production
  - Use the SAME model for document and query embeddings
  - Matryoshka embeddings: can truncate dimensions for speed

Patterns

Document Ingestion Pipeline

import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter';
import { OpenAIEmbeddings }  ;


 splitter =  ({
  : ,
  : ,
  : [, , , , , ],
});

 chunks =  splitter.(documents);


 embeddings =  ({
  : ,
  : ,
});

 vectors =  embeddings.(
  chunks.( chunk.)
);


 ( i = ; i < chunks.; i++) {
   vectorStore.({
    : ,
    : vectors[i],
    : {
      : chunks[i].,
      : chunks[i]..,
      : chunks[i]..,
      : i,
    },
  });
}