📚 AI Knowledge Base — apidcms Documentation

Since version v1.3.25 the AI assistant in the admin panel has access to the entire apidcms documentation. When you ask a question, AI sees not only the database structure but also the full text of all articles from the apidcms-docs repository.

🧠 How It Works

On each request to the AI assistant, the system:

1. Loads the manifest (manifest.json) from the GitHub repo Dezajnisto/apidcms-docs
2. Fetches all articles (36 Markdown files)
3. Strips YAML front-matter — AI receives only the content text
4. Builds the context (~131 KB of text, ~32K tokens)
5. Inserts it into the system prompt before the database data

Result: AI answers questions about apidcms with knowledge of the official documentation.

Example Questions AI Can Now Answer

• "How to create a feedback form with a GDPR consent checkbox?"
• "What hooks are available when developing a plugin?"
• "How does the external page type work?"
• "Explain the template structure: what variables are available in single?"
• "How to configure caching?"
• "What system settings exist and what are they for?"

🗄️ DocsFetcher — Technical Details

Class Core\DocsFetcher

Handles loading and caching of documentation.

$docsFetcher = new \Core\DocsFetcher();
$kb = $docsFetcher->getKnowledgeBase();
// $kb — full documentation text for AI

Caching

The cache directory is created automatically on the first request.

Data Source

• Repository: github.com/Dezajnisto/apidcms-docs (branch master)
• Format: Markdown with YAML front-matter (automatically stripped)
• Source: GitHub Raw (raw.githubusercontent.com)

Fault Tolerance

If GitHub is unavailable:

• Stale cache is used — for up to 6 hours
• If no cache exists, AI works without documentation, with DB context only
• A load error does not break the AI assistant — it continues working

🔄 Cache Invalidation

After updating documentation on GitHub, you can forcibly clear the cache:

$fetcher = new \Core\DocsFetcher();
$count = $fetcher->invalidateCache();
// Removed $count cache files

On the next request, documentation will be loaded fresh.

In the future, invalidation will be available from the admin panel — a "Clear Documentation Cache" button in the AI settings section.

📊 Context Size

32K tokens is ~25% of DeepSeek's context window. Plenty of room remains for conversation and DB data.

🔮 Future Plans

• RAG instead of full load — if the number of articles grows, we'll switch to retrieval-augmented generation: AI will select relevant articles for each specific question
• Admin panel invalidation — a cache-clear button in the UI
• Local mirror — git pull via cron for projects without GitHub access