Vectors WebLLM (#3631)

* Add WebLLM support for vectorization

* Load models when WebLLM extension installed

* Consistency updated

* Move checkWebLlm to initEngine

* Refactor vector request handling to use getAdditionalArgs

* Add error handling for unsupported WebLLM extension

* Add prefix to error causes

public/scripts/extensions.js

@@ -1070,7 +1070,7 @@ export async function installExtension(url, global) {
     toastr.success(t`Extension '${response.display_name}' by ${response.author} (version ${response.version}) has been installed successfully!`, t`Extension installation successful`);
     console.debug(`Extension "${response.display_name}" has been installed successfully at ${response.extensionPath}`);
     await loadExtensionSettings({}, false, false);
-    await eventSource.emit(event_types.EXTENSION_SETTINGS_LOADED);
+    await eventSource.emit(event_types.EXTENSION_SETTINGS_LOADED, response);
 }
 
 /**

public/scripts/extensions/vectors/index.js

@@ -19,6 +19,7 @@ import {
     modules,
     renderExtensionTemplateAsync,
     doExtrasFetch, getApiUrl,
+    openThirdPartyExtensionMenu,
 } from '../../extensions.js';
 import { collapseNewlines, registerDebugFunction } from '../../power-user.js';
 import { SECRET_KEYS, secret_state, writeSecret } from '../../secrets.js';
@@ -34,6 +35,7 @@ import { SlashCommandEnumValue, enumTypes } from '../../slash-commands/SlashComm
 import { slashCommandReturnHelper } from '../../slash-commands/SlashCommandReturnHelper.js';
 import { callGenericPopup, POPUP_RESULT, POPUP_TYPE } from '../../popup.js';
 import { generateWebLlmChatPrompt, isWebLlmSupported } from '../shared.js';
+import { WebLlmVectorProvider } from './webllm.js';
 
 /**
  * @typedef {object} HashedMessage
@@ -60,6 +62,7 @@ const settings = {
     ollama_model: 'mxbai-embed-large',
     ollama_keep: false,
     vllm_model: '',
+    webllm_model: '',
     summarize: false,
     summarize_sent: false,
     summary_source: 'main',
@@ -103,7 +106,7 @@ const settings = {
 };
 
 const moduleWorker = new ModuleWorkerWrapper(synchronizeChat);
-
+const webllmProvider = new WebLlmVectorProvider();
 const cachedSummaries = new Map();
 
 /**
@@ -373,6 +376,8 @@ async function synchronizeChat(batchSize = 5) {
                     return 'Vectorization Source Model is required, but not set.';
                 case 'extras_module_missing':
                     return 'Extras API must provide an "embeddings" module.';
+                case 'webllm_not_supported':
+                    return 'WebLLM extension is not installed or the model is not set.';
                 default:
                     return 'Check server console for more details';
             }
@@ -747,10 +752,11 @@ async function getQueryText(chat, initiator) {
 
 /**
  * Gets common body parameters for vector requests.
- * @returns {object}
+ * @param {object} args Additional arguments
+ * @returns {object} Request body
  */
-function getVectorsRequestBody() {
+function getVectorsRequestBody(args = {}) {
-    const body = {};
+    const body = Object.assign({}, args);
     switch (settings.source) {
         case 'extras':
             body.extrasUrl = extension_settings.apiUrl;
@@ -777,12 +783,30 @@ function getVectorsRequestBody() {
             body.apiUrl = textgenerationwebui_settings.server_urls[textgen_types.VLLM];
             body.model = extension_settings.vectors.vllm_model;
             break;
+        case 'webllm':
+            body.model = extension_settings.vectors.webllm_model;
+            break;
         default:
             break;
     }
     return body;
 }
 
+/**
+ * Gets additional arguments for vector requests.
+ * @param {string[]} items Items to embed
+ * @returns {Promise<object>} Additional arguments
+ */
+async function getAdditionalArgs(items) {
+    const args = {};
+    switch (settings.source) {
+        case 'webllm':
+            args.embeddings = await createWebLlmEmbeddings(items);
+            break;
+    }
+    return args;
+}
+
 /**
  * Gets the saved hashes for a collection
 * @param {string} collectionId
@@ -816,11 +840,12 @@ async function getSavedHashes(collectionId) {
 async function insertVectorItems(collectionId, items) {
     throwIfSourceInvalid();
 
+    const args = await getAdditionalArgs(items.map(x => x.text));
     const response = await fetch('/api/vector/insert', {
         method: 'POST',
         headers: getRequestHeaders(),
         body: JSON.stringify({
-            ...getVectorsRequestBody(),
+            ...getVectorsRequestBody(args),
             collectionId: collectionId,
             items: items,
             source: settings.source,
@@ -858,6 +883,10 @@ function throwIfSourceInvalid() {
     if (settings.source === 'extras' && !modules.includes('embeddings')) {
         throw new Error('Vectors: Embeddings module missing', { cause: 'extras_module_missing' });
     }
+
+    if (settings.source === 'webllm' && (!isWebLlmSupported() || !settings.webllm_model)) {
+        throw new Error('Vectors: WebLLM is not supported', { cause: 'webllm_not_supported' });
+    }
 }
 
 /**
@@ -890,11 +919,12 @@ async function deleteVectorItems(collectionId, hashes) {
  * @returns {Promise<{ hashes: number[], metadata: object[]}>} - Hashes of the results
  */
 async function queryCollection(collectionId, searchText, topK) {
+    const args = await getAdditionalArgs([searchText]);
     const response = await fetch('/api/vector/query', {
         method: 'POST',
         headers: getRequestHeaders(),
         body: JSON.stringify({
-            ...getVectorsRequestBody(),
+            ...getVectorsRequestBody(args),
             collectionId: collectionId,
             searchText: searchText,
             topK: topK,
@@ -919,11 +949,12 @@ async function queryCollection(collectionId, searchText, topK) {
  * @returns {Promise<Record<string, { hashes: number[], metadata: object[] }>>} - Results mapped to collection IDs
  */
 async function queryMultipleCollections(collectionIds, searchText, topK, threshold) {
+    const args = await getAdditionalArgs([searchText]);
     const response = await fetch('/api/vector/query-multi', {
         method: 'POST',
         headers: getRequestHeaders(),
         body: JSON.stringify({
-            ...getVectorsRequestBody(),
+            ...getVectorsRequestBody(args),
             collectionIds: collectionIds,
             searchText: searchText,
             topK: topK,
@@ -1039,6 +1070,72 @@ function toggleSettings() {
     $('#llamacpp_vectorsModel').toggle(settings.source === 'llamacpp');
     $('#vllm_vectorsModel').toggle(settings.source === 'vllm');
     $('#nomicai_apiKey').toggle(settings.source === 'nomicai');
+    $('#webllm_vectorsModel').toggle(settings.source === 'webllm');
+    if (settings.source === 'webllm') {
+        loadWebLlmModels();
+    }
+}
+
+/**
+ * Executes a function with WebLLM error handling.
+ * @param {function(): Promise<T>} func Function to execute
+ * @returns {Promise<T>}
+ * @template T
+ */
+async function executeWithWebLlmErrorHandling(func) {
+    try {
+        return await func();
+    } catch (error) {
+        console.log('Vectors: Failed to load WebLLM models', error);
+        if (!(error instanceof Error)) {
+            return;
+        }
+        switch (error.cause) {
+            case 'webllm-not-available':
+                toastr.warning('WebLLM is not available. Please install the extension.', 'WebLLM not installed');
+                break;
+            case 'webllm-not-updated':
+                toastr.warning('The installed extension version does not support embeddings.', 'WebLLM update required');
+                break;
+        }
+    }
+}
+
+/**
+ * Loads and displays WebLLM models in the settings.
+ * @returns {Promise<void>}
+ */
+function loadWebLlmModels() {
+    return executeWithWebLlmErrorHandling(() => {
+        const models = webllmProvider.getModels();
+        $('#vectors_webllm_model').empty();
+        for (const model of models) {
+            $('#vectors_webllm_model').append($('<option>', { value: model.id, text: model.toString() }));
+        }
+        if (!settings.webllm_model || !models.some(x => x.id === settings.webllm_model)) {
+            if (models.length) {
+                settings.webllm_model = models[0].id;
+            }
+        }
+        $('#vectors_webllm_model').val(settings.webllm_model);
+        return Promise.resolve();
+    });
+}
+
+/**
+ * Creates WebLLM embeddings for a list of items.
+ * @param {string[]} items Items to embed
+ * @returns {Promise<Record<string, number[]>>} Calculated embeddings
+ */
+async function createWebLlmEmbeddings(items) {
+    return executeWithWebLlmErrorHandling(async () => {
+        const embeddings = await webllmProvider.embedTexts(items, settings.webllm_model);
+        const result = /** @type {Record<string, number[]>} */ ({});
+        for (let i = 0; i < items.length; i++) {
+            result[items[i]] = embeddings[i];
+        }
+        return result;
+    });
 }
 
 async function onPurgeClick() {
@@ -1567,6 +1664,30 @@ jQuery(async () => {
         $('#dialogue_popup_input').val(presetModel);
     });
 
+    $('#vectors_webllm_install').on('click', (e) => {
+        e.preventDefault();
+        e.stopPropagation();
+
+        if (Object.hasOwn(SillyTavern, 'llm')) {
+            toastr.info('WebLLM is already installed');
+            return;
+        }
+
+        openThirdPartyExtensionMenu('https://github.com/SillyTavern/Extension-WebLLM');
+    });
+
+    $('#vectors_webllm_model').on('input', () => {
+        settings.webllm_model = String($('#vectors_webllm_model').val());
+        Object.assign(extension_settings.vectors, settings);
+        saveSettingsDebounced();
+    });
+
+    $('#vectors_webllm_load').on('click', async () => {
+        if (!settings.webllm_model) return;
+        await webllmProvider.loadModel(settings.webllm_model);
+        toastr.success('WebLLM model loaded');
+    });
+
     $('#api_key_nomicai').toggleClass('success', !!secret_state[SECRET_KEYS.NOMICAI]);
 
     toggleSettings();
@@ -1578,6 +1699,11 @@ jQuery(async () => {
     eventSource.on(event_types.CHAT_DELETED, purgeVectorIndex);
     eventSource.on(event_types.GROUP_CHAT_DELETED, purgeVectorIndex);
     eventSource.on(event_types.FILE_ATTACHMENT_DELETED, purgeFileVectorIndex);
+    eventSource.on(event_types.EXTENSION_SETTINGS_LOADED, async (manifest) => {
+        if (settings.source === 'webllm' && manifest?.display_name === 'WebLLM') {
+            await loadWebLlmModels();
+        }
+    });
 
     SlashCommandParser.addCommandObject(SlashCommand.fromProps({
         name: 'db-ingest',

public/scripts/extensions/vectors/settings.html

@@ -21,8 +21,24 @@
                     <option value="openai">OpenAI</option>
                     <option value="togetherai">TogetherAI</option>
                     <option value="vllm">vLLM</option>
+                    <option value="webllm" data-i18n="WebLLM Extension">WebLLM Extension</option>
                 </select>
             </div>
+            <div class="flex-container flexFlowColumn" id="webllm_vectorsModel">
+                <label for="vectors_webllm_model" data-i18n="Vectorization Model">
+                    Vectorization Model
+                </label>
+                <div class="flex-container">
+                    <select id="vectors_webllm_model" class="text_pole flex1">
+                    </select>
+                    <div id="vectors_webllm_load" class="menu_button menu_button_icon" title="Verify and load the selected model.">
+                        <i class="fa-solid fa-check-to-slot"></i>
+                    </div>
+                </div>
+                <div>
+                    Requires the WebLLM extension to be installed. Click <a href="#" id="vectors_webllm_install">here</a> to install.
+                </div>
+            </div>
             <div class="flex-container flexFlowColumn" id="ollama_vectorsModel">
                 <label for="vectors_ollama_model" data-i18n="Vectorization Model">
                     Vectorization Model

public/scripts/extensions/vectors/webllm.js

@@ -0,0 +1,64 @@
+export class WebLlmVectorProvider {
+    /** @type {object?} WebLLM engine */
+    #engine = null;
+
+    constructor() {
+        this.#engine = null;
+    }
+
+    /**
+     * Check if WebLLM is available and up-to-date
+     * @throws {Error} If WebLLM is not available or not up-to-date
+     */
+    #checkWebLlm() {
+        if (!Object.hasOwn(SillyTavern, 'llm')) {
+            throw new Error('WebLLM is not available', { cause: 'webllm-not-available' });
+        }
+
+        if (typeof SillyTavern.llm.generateEmbedding !== 'function') {
+            throw new Error('WebLLM is not updated', { cause: 'webllm-not-updated' });
+        }
+    }
+
+    /**
+     * Initialize the engine with a model.
+     * @param {string} modelId Model ID to initialize the engine with
+     * @returns {Promise<void>} Promise that resolves when the engine is initialized
+     */
+    #initEngine(modelId) {
+        this.#checkWebLlm();
+        if (!this.#engine) {
+            this.#engine = SillyTavern.llm.getEngine();
+        }
+
+        return this.#engine.loadModel(modelId);
+    }
+
+    /**
+     * Get available models.
+     * @returns {{id:string, toString: function(): string}[]} Array of available models
+     */
+    getModels() {
+        this.#checkWebLlm();
+        return SillyTavern.llm.getEmbeddingModels();
+    }
+
+    /**
+     * Generate embeddings for a list of texts.
+     * @param {string[]} texts Array of texts to generate embeddings for
+     * @param {string} modelId Model to use for generating embeddings
+     * @returns {Promise<number[][]>} Array of embeddings for each text
+     */
+    async embedTexts(texts, modelId) {
+        await this.#initEngine(modelId);
+        return this.#engine.generateEmbedding(texts);
+    }
+
+    /**
+     * Loads a model into the engine.
+     * @param {string} modelId Model ID to load
+     */
+    async loadModel(modelId) {
+        await this.#initEngine(modelId);
+    }
+}

src/endpoints/vectors.js

@@ -31,6 +31,7 @@ const SOURCES = [
     'ollama',
     'llamacpp',
     'vllm',
+    'webllm',
 ];
 
 /**
@@ -64,6 +65,8 @@ async function getVector(source, sourceSettings, text, isQuery, directories) {
             return getVllmVector(text, sourceSettings.apiUrl, sourceSettings.model, directories);
         case 'ollama':
             return getOllamaVector(text, sourceSettings.apiUrl, sourceSettings.model, sourceSettings.keep, directories);
+        case 'webllm':
+            return sourceSettings.embeddings[text];
     }
 
     throw new Error(`Unknown vector source ${source}`);
@@ -114,6 +117,9 @@ async function getBatchVector(source, sourceSettings, texts, isQuery, directorie
             case 'ollama':
                 results.push(...await getOllamaBatchVector(batch, sourceSettings.apiUrl, sourceSettings.model, sourceSettings.keep, directories));
                 break;
+            case 'webllm':
+                results.push(...texts.map(x => sourceSettings.embeddings[x]));
+                break;
             default:
                 throw new Error(`Unknown vector source ${source}`);
         }
@@ -179,6 +185,11 @@ function getSourceSettings(source, request) {
             return {
                 model: 'nomic-embed-text-v1.5',
             };
+        case 'webllm':
+            return {
+                model: String(request.body.model),
+                embeddings: request.body.embeddings ?? {},
+            };
         default:
             return {};
     }