RAG Intelligence

RAG (Retrieval-Augmented Generation) is the AI technique that powers Keep'em's chat. It's what ensures the AI gives answers based on your actual content rather than generic knowledge.

How RAG Works

When a viewer asks a question, the system follows a specific process:

1. Embed the question — The viewer's question is converted into a vector embedding, a mathematical representation of its meaning.

2. Search for relevant content — The system searches your video transcript chunks and document chunks for the most semantically similar matches. This isn't keyword matching — it understands meaning. "How do I accept payments?" matches content about Stripe integration even if the word "payments" never appears.

3. Retrieve top matches — The most relevant chunks (typically 3–5) are selected based on a similarity threshold. Chunks that don't meet the minimum relevance score are excluded.

4. Generate the response — The AI receives the viewer's question along with the retrieved chunks as context. It generates a natural language response that synthesizes the relevant information.

5. Include sources — The response references where the information came from, including timestamp links back to the relevant point in the video.

Why RAG Matters

Without RAG, an AI chatbot relies on its general training data. It might know that "Stripe is a payment processor" but it has no idea how YOUR product integrates with Stripe, what specific steps YOUR onboarding video shows, or what YOUR FAQ says about supported payment methods.

RAG grounds every response in your specific content. The AI becomes an expert on your product because it's reading your actual documentation and transcript every time it answers.

Confidence and Honesty

The similarity threshold serves as a confidence filter. If the AI can't find sufficiently relevant content to answer a question, it responds honestly rather than guessing. It might say something like: "I don't have specific information about that in the materials I've been provided. Let me connect you with the team."

This honesty is a deliberate design choice. A wrong answer from AI is worse than no answer at all.