Core Concepts

🔄 RAG (Retrieval Augmented Generation)

RAG is a technique that enhances large language models by retrieving relevant information from external knowledge sources before generating a response. This approach allows AI systems to provide more accurate, up-to-date, and contextually relevant answers.

How RAG Works:

1. Retrieval: Search for relevant documents or passages based on the user's query
2. Augmentation: Combine the retrieved context with the original query
3. Generation: Use the enhanced prompt to generate a more informed response

🧮 Vector Embeddings

Vector embeddings are numerical representations of text, images, or other data types in high-dimensional space. Similar content is represented by vectors that are close together in this space, enabling semantic search and similarity matching.

Key Properties:

• Semantic Meaning: Captures the meaning and context of content
• Dimensionality: Typically 384 to 1536 dimensions
• Distance Metrics: Cosine similarity, Euclidean distance, dot product
• Model Dependent: Different embedding models produce different representations

// Example: Text to vector embedding
const text = "Vectorize helps build AI applications";
const embedding = await embed(text);
// Result: [0.1, -0.3, 0.7, 0.2, ...] (1536 dimensions)

🎯 Similarity Search

Similarity search finds the most relevant content by comparing vector embeddings. Instead of exact keyword matching, it finds semantically similar content even when different words are used.

Search Types:

• k-NN Search: Find the k nearest neighbors to a query vector
• Range Search: Find all vectors within a distance threshold
• Filtered Search: Combine similarity with metadata filters
• Hybrid Search: Combine vector search with keyword search

✂️ Chunking Strategies

Chunking is the process of breaking down large documents into smaller, manageable pieces that can be effectively embedded and retrieved. The right chunking strategy is crucial for optimal RAG performance.

Common Chunking Methods:

• Fixed Size: Split by character or token count (e.g., 512 tokens)
• Semantic: Split by meaning, preserving context
• Sentence-based: Split at sentence boundaries
• Recursive: Try larger chunks first, then break down if needed
• Document Structure: Split by headers, paragraphs, or sections

Chunking Considerations:

• Chunk Size: Balance between context and specificity
• Overlap: Include overlapping content to maintain context
• Metadata: Preserve document structure and source information
• Content Type: Different strategies for code, tables, lists

// Example chunking configuration
{
  "strategy": "recursive",
  "chunkSize": 1000,
  "chunkOverlap": 200,
  "separators": ["\n\n", "\n", " ", ""],
  "preserveMetadata": true
}