README

Makes a TYPO3 site readable for AI systems, and gives its visitors a search assistant that answers from the site's own content.

Two halves that work independently:

• Machine-readable content — an llms.txt file following the llmstxt.org specification, and a Markdown representation of every page. Free, no key needed.
• AI search assistant — a chat widget that answers visitor questions from your search index, with links to the pages it used. Requires a subscription key.

Requirements

Installation

composer require web-nomads/wn-ai-bridge

Then activate the extension and flush caches.

Nice URLs

Without route enhancers the endpoints are reachable by page type only. To get readable URLs, import the shipped route configuration:

# config/sites/<identifier>/config.yaml
imports:
  -
    resource: 'EXT:wn_ai_bridge/Configuration/Routes/RouterEnhancer.yaml'

So https://example.com/about also exists as https://example.com/about.md.

What llms.txt is for

llms.txt sits at a well-known location and tells language models what a site is about: a short description, its structure, and links to machine-readable versions of the content. The place is the same idea as robots.txt, the purpose is not — it describes content rather than restricting access. This extension generates it from your actual page tree, so it stays correct without anyone maintaining it by hand.

Configure the metadata (topics, contact, description) per site on the AI Bridge tab of the site configuration.

AI search assistant

A floating chat widget. Switch it on in the extension configuration (assistantEnabled) and per site in the site configuration.

Search-only (no API key) returns ranked matching pages as suggestions with links. Fast, free, and nothing leaves your server.

Hybrid (with an LLM API key in assistantApiKey) additionally lets the model compose a short answer from the retrieved pages and cite them. Any failure — quota, timeout, malformed response — falls back to search-only rather than showing an error.

The assistant reads ke_search and indexed_search when they are installed, and always keeps a dependency-free pages/tt_content fallback so it returns something even without a search index.

Backend modules

Answers is the local learning source. An entry is played back verbatim when a new question matches it in meaning — term overlap plus string similarity, not exact wording. Weaker matches are handed to the model as binding hints. Entries come from three places: written by an editor, taken over from a logged answer in Enquiries, or captured from a correction a visitor made in the chat, which arrives as "pending" and is only used once approved.

Subscription

The assistant and its two modules are subscription features. Enter the key in the Subscription key field of the extension configuration. The key is encrypted and signed; it carries the domains it is valid for, an expiry date and the enabled features.

Important:
Without a valid key the widget stays hidden and the two modules disappear. Everything else keeps working.

Order your subscription key:
DE: https://www.marcelmarty.ch/#extensions
EN: https://www.marcelmarty.ch/en/#extensions

Email me for a 14-day free trial licence.

What the licence check sends

Once a day the extension asks the issuing server whether the subscription is still active, sending the subscription id, this installation's hostname and a random nonce. No visitor data is involved — no IP addresses, no questions, no page content. The signed answer is what carries a renewal to the installation, so a renewed subscription takes effect without anyone pasting a new key, and a revoked one stops working without waiting for its expiry date.

An unreachable server changes nothing: the date inside the key decides, and only an explicitly signed "revoked" switches the features off.

Leave subscriptionKey empty and nothing is ever sent.

See Documentation/Administrator for the full description, including what is reported when an installation looks manipulated.

Configuration

Extension configuration (Admin Tools → Settings → Extension Configuration) covers the assistant, the LLM provider, rate limiting and the subscription key. Per-site settings live on the AI Bridge and AI Search Assistant tabs of the site configuration.

Two things worth setting before going live with the assistant:

• Enable the rate limiter (rateLimiterEnabled). The assistant endpoint is reachable without authentication and every request can cost money.
• Set a spending limit in your LLM provider account as an independent second net.

Documentation

The full manual is rendered at docs.typo3.org, and its source lives in Documentation/.

Development

composer install

composer test        # unit tests
composer stan        # PHPStan level 6
composer cs:check    # coding standards, --dry-run
composer cs:fix      # apply them
composer ci          # all of the above

composer release     # build the TER archive

composer release writes wn_ai_bridge_<version>.zip next to the extension folder and refuses to build if the version in ext_emconf.php and composer.json disagree, if ext_emconf.php would not land at the archive root, or if anything generated would be packed.

Contributing

Issues and pull requests are welcome at github.com/web-nomads/wn-ai-bridge.

For a pull request: follow the TYPO3 coding standards (composer cs:fix), add tests for behaviour you change, and keep composer ci green.

Credits

This extension started as a fork of web-vision/ai-llms-txt by web-vision, which provides the llms.txt generation according to the llmstxt.org specification. The AI search assistant, the Markdown endpoints and the subscription handling were added here.

Licence

GPL-2.0-or-later, the same licence as the original — see LICENSE.