How Much Data Do You Need to Train an AI Chatbot

Start Small and Expand

The most effective approach is to start with the content that covers your most common customer questions, then add more over time based on what the chatbot struggles with. Most businesses find that 5 to 10 well-written pages of content cover 80% of the questions they receive. The remaining 20% can be added as you identify gaps.

A typical starting point looks like this:

• Your FAQ page or a list of 20 to 30 common questions with answers
• Your main product or service descriptions
• Your return, shipping, and contact policies
• Your pricing information

This amount of content usually produces 50 to 150 chunks, costs 150 to 450 credits ($0.15 to $0.45) to embed, and gives the chatbot enough knowledge to handle the majority of visitor questions.

Data Quantity Guidelines by Use Case

Simple FAQ Chatbot

For a chatbot that answers basic questions about your business, 1 to 5 pages of content is enough. A well-organized FAQ document with 30 to 50 question-answer pairs covers most small business needs. This produces roughly 30 to 100 chunks.

Product Support Chatbot

For a chatbot that handles detailed product questions, you need product descriptions, specifications, troubleshooting guides, and compatibility information. Typically 10 to 50 pages of content, producing 200 to 1,000 chunks. The more detailed your product documentation, the more specific and accurate the chatbot's answers will be.

Internal Knowledge Base

For an internal chatbot that helps employees find company information, you might upload entire policy manuals, training materials, and process documents. This could be 50 to 200+ pages, producing 1,000 to 5,000+ chunks. The cost scales linearly at 3 credits per chunk.

Comprehensive Customer Service

For a full customer service chatbot that handles complex inquiries across products, policies, and procedures, plan for 20 to 100 pages of well-organized content. Include your support team's standard responses to common scenarios. This typically produces 500 to 2,500 chunks.

When More Data Helps

Adding more data helps when customers are asking questions your chatbot cannot answer, or when answers are too vague because the source content lacks detail. The solution is always to add specific, detailed content about the topic the chatbot is struggling with, not to dump in large volumes of loosely related material.

For example, if customers keep asking about sizing and your chatbot gives generic answers, uploading a detailed sizing chart with measurement instructions will immediately improve those responses. You do not need to re-upload everything else.

When More Data Hurts

Adding low-quality, redundant, or contradictory data can actually make your chatbot worse. Common problems include:

• Duplicate content: Uploading the same information in multiple documents means the retrieval system pulls in redundant chunks, wasting context window space that could be used for additional relevant information
• Contradictory information: If one document says your return window is 30 days and another says 60 days, the chatbot may cite either one randomly. Clean up contradictions before uploading. See What Happens When Training Data Contradicts Itself.
• Outdated content: Old pricing, discontinued products, or former policies will surface in search results and produce incorrect answers
• Off-topic content: Adding content that is unrelated to what customers ask about adds noise to the search results without improving answers

Cost Scaling

Embedding costs are predictable and linear. At 3 credits per chunk:

• Small FAQ (50 chunks): 150 credits ($0.15)
• Medium knowledge base (500 chunks): 1,500 credits ($1.50)
• Large documentation library (2,000 chunks): 6,000 credits ($6.00)
• Enterprise scale (10,000 chunks): 30,000 credits ($30.00)

These are one-time costs. Once embedded, there is no ongoing storage charge for your training data. You only pay again if you upload new content or re-embed existing content.