>>> [!note] Migrated issue <!-- Drupal.org comment --> <!-- Migrated from issue #3616670. --> Reported by: [joshua1234511](https://www.drupal.org/user/3362218) Related to !19 >>> <h3>Problem/Motivation</h3> <p><code>EmpathyEvaluator::evaluateScenario()</code> builds its AI request tags<br> internally in <code>buildRequestTags()</code> and passes them to<br> <code>$provider-&amp;gt;chat($input, $model_id, $tags)</code>. There is no parameter<br> and no alter hook, so a calling module cannot add a tag to the request.</p> <p>This matters because provider plugins use tags as their control channel. In<br> <a href="https://www.drupal.org/project/ai_provider_openai">AI Provider OpenAI</a>,<br> <code>OpenAiProvider::moderationEndpoints()</code> skips the OpenAI moderation<br> call when the request carries a <code>skip_moderation</code> tag:</p> <pre><div class="codeblock"><code class="language-php">if (!$this-&amp;gt;moderation || in_array('skip_moderation', $tags)) {<br>&nbsp; return;<br>}<p>We call <code>evaluateScenario()</code> over our own authored clinical and HR<br> ethics scenarios to build a baseline of model answers for comparison. Several are<br> refused with:</p> <blockquote><p> </p><pre>AiUnsafePromptException: The prompt was flagged by the<br>moderation model.</pre> </blockquote> <p>The refused scenarios cover triage under scarcity, a safeguarding disclosure<br> by a minor, and reporting duties for assault injuries. These are legitimate<br> professional training material, authored and reviewed by the site owner. They<br> are not untrusted user input.</p> <p>Because the tag cannot be injected, the only ways to get those scenarios<br> evaluated are:</p> <ol> <li>Set <code>ai_provider_openai.settings:moderation</code> to <code>FALSE</code>,<br> which disables moderation for <em>every</em> OpenAI call on the site &mdash; including<br> calls that score free text submitted by anonymous visitors. That is precisely the<br> input moderation should apply to, so this trades away the wrong risk.</li> <li>Patch AI Empathy locally and carry the patch.</li> </ol> <p>Neither is a good option for a site that wants moderation on untrusted input<br> and off for its own vetted corpus.</p> <h3>Proposed resolution</h3> <p>Add an optional <code>$additional_tags</code> parameter to<br> <code>evaluateScenario()</code>, merged with the tags it already builds:</p> <pre><div class="codeblock"><code class="language-php">public function evaluateScenario(<br>&nbsp; AiEmpathyScenarioInterface $scenario,<br>&nbsp; ?string $provider_model_option = NULL,<br>&nbsp; int $run_number = 1,<br>&nbsp; ?int $session_id = NULL,<br>&nbsp; ?string $additional_context = NULL,<br>&nbsp; ?string $benchmark_id = NULL,<br>&nbsp; array $additional_tags = [],<br>): AiEmpathyResult {<br>&nbsp; // ...<br>&nbsp; $tags = array_merge(<br>&nbsp;&nbsp;&nbsp; $this-&amp;gt;buildRequestTags($scenario, $run_number),<br>&nbsp;&nbsp;&nbsp; $additional_tags<br>&nbsp; );<br>}<p>Appending is backwards compatible: the parameter is optional and defaults to<br> an empty array, so existing callers are unaffected and the tags AI Empathy relies<br> on for its own observability are still present and still first.</p> <p>An alter hook or event on the tag list would also solve this and would be more<br> flexible for other use cases. The parameter is proposed instead because it is<br> smaller, and because it does not require a caller to implement an event<br> subscriber in order to add one string.</p> <h3>Remaining tasks</h3> <ul> <li>Add the parameter and merge it into the tag list</li> <li>Accept and forward the same parameter from <code>evaluateScenarioBatch()</code></li> <li>Document that provider plugins may treat tags as control flags, with<br> <code>skip_moderation</code> as the worked example</li> <li>Test coverage asserting caller-supplied tags reach the provider</li> </ul> <h3>User interface changes</h3> <p>None.</p> <h3>API changes</h3> <p>None breaking. One new optional parameter with a default value, appended to<br> the existing signature.</p> <h3>Data model changes</h3> <p>None.</p> </code></div></pre></code></div></pre>