Problem

Human readers can click hyperlinks to learn more. AI models often process documentation in isolated “chunks”. When an AI encounters a hyperlink, it cannot “click” it to fetch context. If critical information is hidden behind a link, the AI may fail to provide accurate answers. This leads to hallucinations.

Why it matters

AI models rely on immediate surrounding text to determine meaning. If your documentation is highly fragmented with poor context preservation, AI-driven search tools (like RAG systems) will struggle to provide high-quality responses.

Approach

Use Inline Context Unrolling to provide the minimum viable context alongside every major link. Use docmd’s LLMs Plugin to provide a unified, machine-readable view of your entire documentation set.

Implementation

1. Descriptive Linking and Summaries

Avoid generic link text. Provide a brief, one-sentence summary of the linked concept alongside the link.

• ❌ Poor (Context Lost): To configure the timeout, refer to the General Configuration.
• ✅ Better (Context Preserved): You can configure the timeoutMs parameter within the General Configuration, which defines how long the engine waits before failing a network request.

2. Using Collapsible Sections for Detail

Collapsible Containers are excellent for AI optimisation. The content remains part of the raw Markdown source for the AI, but it is visually tucked away for human readers.

### Database Connection

Connect using the primary URI.

::: collapsible "What is the URI format?"
The URI follows the standard PostgreSQL format: `postgresql://user:password@host:port/database`.
:::

3. Enabling the LLMs Plugin

Enable the LLMs Plugin in your docmd.config.json. This plugin generates a llms-full.txt file after every build. It concatenates your entire documentation set into a single, high-context file that LLMs consume easily.

Trade-offs

Inline context unrolling makes documentation slightly more verbose and introduces minor redundancy. However, this is a small price to ensure your documentation is “AI-ready” and capable of powering high-quality automated support.