Best Practices

The system prompt (instructions) is the most important factor in shaping your AI agent's behavior and personality. Clear, specific instructions ensure your agent responds consistently and aligns with your brand voice and business goals.

A well-crafted system prompt should define the agent's role, tone, constraints, and formatting preferences. Use the examples below as templates and customize them for your specific use case.

The quality of your AI agent's responses depends heavily on the quality and structure of your data sources. Well-organized, clearly formatted documents with consistent headings and bullet points improve retrieval accuracy and response quality.

Avoid long walls of text, ambiguous pronouns, and complex nested structures. Instead, use clear headings, concise sentences, and explicit key-value pairs for factual information like pricing, policies, and specifications.

For your AI agent to suggest specific URLs in responses, those links must be explicitly included in your training data. While web crawling extracts content for context, the agent may not remember exact URLs unless they're clearly documented.

Create a dedicated document that maps page names to URLs. This ensures your agent can provide accurate links when users ask for resources, documentation pages, or specific content.

Example Link Mapping Document:

Docimal agents can display images in chat responses using standard markdown image syntax: ![Alt Text](image-url)

To enable image display, include instructions in your system prompt that tell the agent when and how to show images. Ensure image URLs are publicly accessible and end with common image extensions (.png, .jpg, .gif, .webp).

Example System Prompt Addition:

Selecting the right AI model is crucial for balancing speed, cost, accuracy, and reasoning capability. Docimal supports multiple model providers and tiers, allowing you to optimize for your specific use case and budget.

Available Models: GPT-4o, GPT-4, GPT-4 Turbo, GPT-4o mini, Claude 3.5 Sonnet, Claude 3 Opus, Claude 3 Haiku, Gemini 1.5 Pro, and more.

Even with well-structured documents, your AI agent may occasionally provide suboptimal answers or miss key nuances. Docimal's Revise feature and Q&A Data tool allow you to manually correct responses and teach your agent specific answers for common edge cases.

Revise Feature: Review past conversations in the Activity tab, identify incorrect or incomplete responses, and use the "Revise" button to provide the correct answer. The agent learns from these corrections over time.

Q&A Data Type: Upload explicit question-answer pairs in the Knowledge tab. These are prioritized over general document retrieval, ensuring your agent gives exact answers for critical questions like pricing, policies, and key features.