Skip to main content

What is RAG?

RAG (Retrieval-Augmented Generation) is a technique that combines semantic search with AI text generation. Instead of the model making up answers, it consults your real knowledge base and generates grounded answers.

How It Works in Tartini

1. Knowledge Base

Your documents, FAQs, and website content are processed and stored in a vector database (embeddings). Each piece of content becomes a “chunk” with metadata. When you send a query, the system converts the text into a vector and searches for the most similar chunks, even if they use different words.
“How do I return a product?” finds content about “return policy” even without using the same words.

3. Answer Generation (optional)

If you use returnMode: "ai_generated_answer", an AI model generates a natural answer using the chunks found as its basis.

Key Concepts

Source Types

Audience

Score Threshold

The scoreThreshold (0 to 1) defines the minimum relevance level:
  • 0.5 - 0.6: More results, lower precision
  • 0.7 - 0.8: Balance between quantity and quality (recommended)
  • 0.9+: Only highly relevant results

Reranking

When useRerank: true, results go through a second model that reorders them by actual relevance.

Return Modes

chunks_only

Returns only the raw chunks. Useful when you want to process the results in your own system.

ai_generated_answer

Returns the chunks and an AI-generated answer.
The ai_generated_answer mode is slower because it includes the generation step. Use chunks_only when latency is critical.

Related topics

RAG Retrieve