Customizing AI Behavior

AI Docs provides powerful capabilities to automate documentation, but it also offers extensive flexibility to tailor its AI behavior to your specific needs. This includes fine-tuning how the AI generates content, how it responds in chat, and even how its outputs are structured for consumption by other AI systems.

Tailoring AI Generation and RAG Behavior

You can directly influence the core AI processes within AI Docs, from content generation to the interactive AI chat assistant.

Prompt Engineering for RAG Chat

The Retrieval Augmented Generation (RAG) chat assistant uses a systemPrompt to guide its persona, instructions, and how it synthesizes information. Customizing this prompt allows you to refine the AI's responses and interaction style.

By modifying the systemPrompt within the RAG chat endpoint configuration, you can:

• Adjust Persona: Make the AI more formal, informal, concise, or verbose.
• Refine Instructions: Provide specific guidelines on how to answer questions, what to prioritize, or how to handle ambiguous queries.
• Control Context Usage: Instruct the AI on how to best utilize the retrieved code and documentation context, for example, by emphasizing documentation over raw code for conceptual answers.

Swapping AI Models

AI Docs is designed to be model-agnostic, allowing you to integrate different Large Language Models (LLMs) and embedding models. This flexibility enables you to experiment with various models to find the best fit for your project's needs in terms of performance, cost, and output quality.

• Embedding Models: For semantic understanding and vector generation, AI Docs primarily uses OpenAI's text-embedding-3-small. You can swap this for other compatible embedding models by modifying the relevant streamText calls and ensuring the necessary API keys are configured.
• Content Generation Models: For generating MDX documentation content, AI Docs currently leverages Google's gemini-2.5-flash.
• RAG Chat Models: The interactive AI chat assistant is powered by OpenAI's gpt-4o-mini.

To swap models, you would typically:

1. Configure API Keys: Ensure you have the API keys for your chosen model provider (e.g., Google Gemini API key) set up in your Environment Configuration.
2. Modify Model Calls: Update the model parameter in the streamText calls within the AI generation and RAG chat logic to point to your desired model.

Refining Retrieval Augmented Generation (RAG)

The effectiveness of the AI chat assistant heavily relies on its ability to retrieve relevant context. You can customize the RAG pipeline to optimize how information is found and presented to the LLM.

• Retrieval Strategy: The searchRelevantChunks function determines how relevant code and documentation chunks are retrieved from Vector Search with Qdrant. You can modify parameters like topK (number of chunks to retrieve) or implement custom re-ranking logic to improve the quality of retrieved context.
• Context Assembly: The assembleContext function is responsible for structuring the retrieved chunks into a coherent input for the LLM. You can adjust how documentation snippets are prioritized over code, how chunks are formatted, or even introduce additional filtering to ensure the LLM receives the most pertinent information.

Configuring AI-Ready Documentation Outputs

AI Docs generates documentation that is not only human-readable but also specifically formatted for consumption by other AI systems. These specialized endpoints allow you to integrate your project's knowledge base into custom LLM workflows, chatbots, or other automated tools.

The llms-full.txt Endpoint

This endpoint provides the complete, concatenated content of all your generated documentation pages in a single plain text file. It's ideal for scenarios where an external AI needs a comprehensive, raw dump of your project's knowledge base for deep analysis, fine-tuning, or extensive context.

• Purpose: Offers a full-text index of your project's documentation.
• Content: Includes all titles, groups, and the full markdown content of each page, separated by clear delimiters.
• Use Cases: Training custom LLMs, creating a comprehensive knowledge base for internal AI tools, or performing advanced text analysis.

The llms.txt Endpoint

This endpoint provides a structured, summarized index of your documentation, designed for quick semantic lookup and integration into RAG systems. Each entry includes the page title, a direct URL, and a concise one-line description extracted from the page's content.

• Purpose: Serves as a lightweight, semantic index for RAG systems or other LLMs needing quick overviews.
• Content: A list of [Page Title](URL): One-line description. The description is automatically extracted from the beginning of each page's content, stripping markdown and code blocks.
• Use Cases: Providing context to external chatbots, building custom search interfaces, or feeding a concise overview of documentation pages to an LLM.

The skill.md Endpoint

This endpoint generates a Markdown file that acts as a comprehensive overview of your project, including its description, repository links, and a table of contents for all documentation pages. It's designed to give an external LLM a high-level understanding of your project's "skills" and available documentation.

• Purpose: Provides a structured project summary for LLMs to understand project capabilities and documentation structure.
• Content: Includes project name, description (from GitHub or project settings), repository URL, links to llms.txt and llms-full.txt, and a grouped table of contents for all documentation pages. It also highlights "Quick Start" pages based on the first few documents.
• Use Cases: Enabling an LLM to answer high-level questions about your project, generating project summaries, or helping an AI navigate your documentation.

Influencing AI-Ready Outputs

While the structure of these endpoints is fixed, you can significantly influence the content they serve:

• Primary Documentation Generation: By customizing the AI's behavior during the initial documentation generation (e.g., through prompt engineering for content creation), you directly affect the doc.content that populates these endpoints.
  • For llms.txt, ensure your generated documentation pages start with a clear, concise introductory sentence to be effectively captured as the one-line description.
• Project Settings: The skill.md endpoint uses the project's name, description, and repository URL. You can update these details in Project Settings and Customization.
• Documentation Page Management: The order and grouping of your documentation pages, managed in Managing Documentation Pages, directly impact the structure and "Quick Start" section of skill.md, as well as the overall order in llms-full.txt and llms.txt.

Best Practices for Customization

When customizing AI behavior in AI Docs, consider these best practices:

• Iterative Testing: Any changes to prompts, models, or RAG logic should be thoroughly tested to ensure they produce accurate, relevant, and desired outputs.
• Cost Management: Be mindful of the cost implications when experimenting with different AI models or increasing parameters like topK in RAG, as these can affect API usage.
• Security: Always validate and sanitize inputs and outputs, especially when integrating with external LLMs, to prevent prompt injection vulnerabilities or unintended data exposure.
• Refer to Documentation: For common issues or deeper insights into the AI/ML core, consult Troubleshooting and Support and AI/ML Core: OpenAI & LangChain.