Expressions: Classify using LLM

Rather than using a separate BERT model to classify the last message,
use the LLM itself to get the classified expression label as a JSON
and set that as the current sprite. Doing this should take more information
into consideration and cut down on extra processing.

This is made possible by the use of constrained generation with JSON
schemas. Only available to TabbyAPI since it's the only backend that
supports the use of JSON schemas, but there can hopefully be a way
to use this with other backends as well.

Intercepts the generation and sets top_k = 1 (for greedy sampling)
and the json_schema to an emotion enum. Doing this also prevents
reingestion of the entire context every time a message is sent and
then asked to be classified, which doesn't compromise the chat
experience.

Signed-off-by: kingbri <bdashore3@proton.me>
This commit is contained in:
kingbri
2024-04-12 01:49:09 -04:00
parent 626c93a1ab
commit 6b656bf380
2 changed files with 122 additions and 58 deletions

View File

@@ -6,10 +6,6 @@
</div>
<div class="inline-drawer-content">
<label class="checkbox_label" for="expression_local" title="Use classification model without the Extras server.">
<input id="expression_local" type="checkbox" />
<span data-i18n="Local server classification">Local server classification</span>
</label>
<label class="checkbox_label" for="expression_translate" title="Use the selected API from Chat Translation extension settings.">
<input id="expression_translate" type="checkbox">
<span>Translate text to English before classification</span>
@@ -22,6 +18,16 @@
<input id="image_type_toggle" type="checkbox">
<span>Image Type - talkinghead (extras)</span>
</label>
<div class="expression_api_block m-b-1 m-t-1">
<label for="expression_api">Classifier API</label>
<small>Select the API for classifying expressions.</small>
<select id="expression_api" class="flex1 margin0" data-i18n="Expression API" placeholder="Expression API">
<option value="0">Local</option>
<option value="1">Extras</option>
<option value="2">LLM</option>
<option value="3">TalkingHead</option>
</select>
</div>
<div class="expression_fallback_block m-b-1 m-t-1">
<label for="expression_fallback">Default / Fallback Expression</label>
<small>Set the default and fallback expression being used when no matching expression is found.</small>