@elizaos/plugin-local-ai 1.0.0-beta.26 → 1.0.0-beta.28

package/dist/index.js

@@ -5,7 +5,6 @@ import path5 from "node:path";
 import { Readable as Readable2 } from "node:stream";
 import { fileURLToPath } from "node:url";
 import { ModelType as ModelType2, logger as logger9 } from "@elizaos/core";
-import pkg from "fastembed";
 import {
   LlamaChatSession,
   getLlama as getLlama2
@@ -21,7 +20,7 @@ var configSchema = z.object({
   STUDIOLM_SERVER_URL: z.string().default("http://localhost:1234"),
   STUDIOLM_SMALL_MODEL: z.string().default("lmstudio-community/deepseek-r1-distill-qwen-1.5b"),
   STUDIOLM_MEDIUM_MODEL: z.string().default("deepseek-r1-distill-qwen-7b"),
-  STUDIOLM_EMBEDDING_MODEL: z.union([z.boolean(), z.string()]).default(false)
+  STUDIOLM_EMBEDDING_MODEL: z.string().default("BAAI/bge-small-en-v1.5")
 });
 function validateModelConfig(config) {
   logger.info("Validating model configuration with values:", {
@@ -47,7 +46,7 @@ async function validateConfig(config) {
       STUDIOLM_SERVER_URL: config.STUDIOLM_SERVER_URL || "http://localhost:1234",
       STUDIOLM_SMALL_MODEL: config.STUDIOLM_SMALL_MODEL || "lmstudio-community/deepseek-r1-distill-qwen-1.5b",
       STUDIOLM_MEDIUM_MODEL: config.STUDIOLM_MEDIUM_MODEL || "deepseek-r1-distill-qwen-7b",
-      STUDIOLM_EMBEDDING_MODEL: config.STUDIOLM_EMBEDDING_MODEL || false
+      STUDIOLM_EMBEDDING_MODEL: config.STUDIOLM_EMBEDDING_MODEL || "BAAI/bge-small-en-v1.5"
     };
     const validatedConfig = configSchema.parse(fullConfig);
     return validatedConfig;
@@ -90,6 +89,18 @@ var MODEL_SPECS = {
       type: "llama"
     }
   },
+  embedding: {
+    name: "bge-small-en-v1.5.Q4_K_M.gguf",
+    repo: "ChristianAzinn/bge-small-en-v1.5-gguf",
+    size: "133 MB",
+    quantization: "Q4_K_M",
+    contextSize: 512,
+    dimensions: 384,
+    tokenizer: {
+      name: "ChristianAzinn/bge-small-en-v1.5-gguf",
+      type: "llama"
+    }
+  },
   vision: {
     name: "Florence-2-base-ft",
     repo: "onnx-community/Florence-2-base-ft",
@@ -2341,7 +2352,6 @@ var VisionManager = class _VisionManager {
 };
 
 // src/index.ts
-var { EmbeddingModel, FlagEmbedding } = pkg;
 var __filename = fileURLToPath(import.meta.url);
 var __dirname = path5.dirname(__filename);
 var wordsToPunish = [
@@ -2399,17 +2409,20 @@ var LocalAIManager = class _LocalAIManager {
   llama;
   smallModel;
   mediumModel;
+  embeddingModel;
+  embeddingContext;
   ctx;
   sequence;
   chatSession;
   modelPath;
   mediumModelPath;
+  embeddingModelPath;
   cacheDir;
-  embeddingModel = null;
   tokenizerManager;
   downloadManager;
   visionManager;
   activeModelConfig;
+  embeddingModelConfig;
   transcribeManager;
   ttsManager;
   studioLMManager;
@@ -2447,6 +2460,7 @@ var LocalAIManager = class _LocalAIManager {
     }
     this.modelPath = path5.join(this.modelsDir, "DeepHermes-3-Llama-3-3B-Preview-q4.gguf");
     this.mediumModelPath = path5.join(this.modelsDir, "DeepHermes-3-Llama-3-8B-q4.gguf");
+    this.embeddingModelPath = path5.join(this.modelsDir, "bge-small-en-v1.5.Q4_K_M.gguf");
     const cacheDirEnv = process.env.CACHE_DIR?.trim();
     if (cacheDirEnv) {
       this.cacheDir = path5.resolve(cacheDirEnv);
@@ -2467,6 +2481,7 @@ var LocalAIManager = class _LocalAIManager {
       this.studioLMManager = StudioLMManager.getInstance();
     }
     this.activeModelConfig = MODEL_SPECS.small;
+    this.embeddingModelConfig = MODEL_SPECS.embedding;
   }
   /**
    * Retrieves the singleton instance of LocalAIManager. If an instance does not already exist, a new one is created and returned.
@@ -2531,18 +2546,31 @@ var LocalAIManager = class _LocalAIManager {
   }
   /**
    * Downloads the model based on the modelPath provided.
-   * Determines whether to download a large or small model based on the current modelPath.
+   * Determines the model spec and path based on the model type.
    *
+   * @param {ModelTypeName} modelType - The type of model to download
+   * @param {ModelSpec} [customModelSpec] - Optional custom model spec to use instead of the default
    * @returns A Promise that resolves to a boolean indicating whether the model download was successful.
    */
-  async downloadModel(modelType) {
-    const modelSpec = modelType === ModelType2.TEXT_LARGE ? MODEL_SPECS.medium : MODEL_SPECS.small;
-    const modelPath = modelType === ModelType2.TEXT_LARGE ? this.mediumModelPath : this.modelPath;
+  async downloadModel(modelType, customModelSpec) {
+    let modelSpec;
+    let modelPath;
+    if (customModelSpec) {
+      modelSpec = customModelSpec;
+      modelPath = modelType === ModelType2.TEXT_EMBEDDING ? this.embeddingModelPath : modelType === ModelType2.TEXT_LARGE ? this.mediumModelPath : this.modelPath;
+    } else if (modelType === ModelType2.TEXT_EMBEDDING) {
+      modelSpec = MODEL_SPECS.embedding;
+      modelPath = this.embeddingModelPath;
+    } else {
+      modelSpec = modelType === ModelType2.TEXT_LARGE ? MODEL_SPECS.medium : MODEL_SPECS.small;
+      modelPath = modelType === ModelType2.TEXT_LARGE ? this.mediumModelPath : this.modelPath;
+    }
     try {
       return await this.downloadManager.downloadModel(modelSpec, modelPath);
     } catch (error) {
       logger9.error("Model download failed:", {
         error: error instanceof Error ? error.message : String(error),
+        modelType,
         modelPath
       });
       throw error;
@@ -2594,31 +2622,107 @@ var LocalAIManager = class _LocalAIManager {
         logger9.warn("Models directory does not exist, creating it:", this.modelsDir);
         fs5.mkdirSync(this.modelsDir, { recursive: true });
       }
+      await this.downloadModel(ModelType2.TEXT_EMBEDDING);
+      if (!this.llama) {
+        this.llama = await getLlama2();
+      }
       if (!this.embeddingModel) {
-        logger9.info("Creating new FlagEmbedding instance with BGESmallENV15 model");
-        const barLength = 30;
-        const emptyBar = "\u25B1".repeat(barLength);
-        logger9.info(`Downloading embedding model: ${emptyBar} 0%`);
-        this.embeddingModel = await FlagEmbedding.init({
-          cacheDir: this.modelsDir,
-          model: EmbeddingModel.BGESmallENV15,
-          maxLength: 512,
-          showDownloadProgress: true
+        logger9.info("Loading embedding model:", this.embeddingModelPath);
+        this.embeddingModel = await this.llama.loadModel({
+          modelPath: this.embeddingModelPath,
+          gpuLayers: 0,
+          // Embedding models are typically small enough to run on CPU
+          vocabOnly: false
+        });
+        this.embeddingContext = await this.embeddingModel.createEmbeddingContext({
+          contextSize: this.embeddingModelConfig.contextSize,
+          batchSize: 512
         });
-        const completedBar = "\u25B0".repeat(barLength);
-        logger9.info(`Downloading embedding model: ${completedBar} 100%`);
-        logger9.success("FlagEmbedding instance created successfully");
+        logger9.success("Embedding model initialized successfully");
       }
     } catch (error) {
       logger9.error("Embedding initialization failed with details:", {
         error: error instanceof Error ? error.message : String(error),
         stack: error instanceof Error ? error.stack : void 0,
         modelsDir: this.modelsDir,
-        model: EmbeddingModel.BGESmallENV15
+        embeddingModelPath: this.embeddingModelPath
       });
       throw error;
     }
   }
+  /**
+   * Generate embeddings using the proper LlamaContext.getEmbedding method.
+   */
+  async generateEmbedding(text) {
+    try {
+      await this.lazyInitEmbedding();
+      if (!this.embeddingModel || !this.embeddingContext) {
+        throw new Error("Failed to initialize embedding model");
+      }
+      logger9.info("Generating embedding for text", { textLength: text.length });
+      const embeddingResult = await this.embeddingContext.getEmbeddingFor(text);
+      const mutableEmbedding = [...embeddingResult.vector];
+      const normalizedEmbedding = this.normalizeEmbedding(mutableEmbedding);
+      logger9.info("Embedding generation complete", { dimensions: normalizedEmbedding.length });
+      return normalizedEmbedding;
+    } catch (error) {
+      logger9.error("Embedding generation failed:", {
+        error: error instanceof Error ? error.message : String(error),
+        stack: error instanceof Error ? error.stack : void 0,
+        textLength: text?.length ?? "text is null"
+      });
+      const zeroDimensions = process.env.LOCAL_EMBEDDING_DIMENSIONS ? parseInt(process.env.LOCAL_EMBEDDING_DIMENSIONS, 10) : this.embeddingModelConfig.dimensions;
+      return new Array(zeroDimensions).fill(0);
+    }
+  }
+  /**
+   * Normalizes an embedding vector using L2 normalization
+   *
+   * @param {number[]} embedding - The embedding vector to normalize
+   * @returns {number[]} - The normalized embedding vector
+   */
+  normalizeEmbedding(embedding) {
+    const squareSum = embedding.reduce((sum, val) => sum + val * val, 0);
+    const norm = Math.sqrt(squareSum);
+    if (norm === 0) {
+      return embedding;
+    }
+    return embedding.map((val) => val / norm);
+  }
+  /**
+   * Lazy initialize the embedding model
+   */
+  async lazyInitEmbedding() {
+    if (this.embeddingInitialized) return;
+    if (!this.embeddingInitializingPromise) {
+      this.embeddingInitializingPromise = (async () => {
+        try {
+          await this.initializeEnvironment();
+          await this.downloadModel(ModelType2.TEXT_EMBEDDING);
+          if (!this.llama) {
+            this.llama = await getLlama2();
+          }
+          this.embeddingModel = await this.llama.loadModel({
+            modelPath: this.embeddingModelPath,
+            gpuLayers: 0,
+            // Embedding models are typically small enough to run on CPU
+            vocabOnly: false
+          });
+          this.embeddingContext = await this.embeddingModel.createEmbeddingContext({
+            contextSize: this.embeddingModelConfig.contextSize,
+            batchSize: 512
+          });
+          this.embeddingInitialized = true;
+          logger9.info("Embedding model initialized successfully");
+        } catch (error) {
+          logger9.error("Failed to initialize embedding model:", error);
+          this.embeddingInitializingPromise = null;
+          throw error;
+        }
+      })();
+    }
+    await this.embeddingInitializingPromise;
+  }
   /**
    * Asynchronously generates text using StudioLM models based on the specified parameters.
    *
@@ -2740,30 +2844,6 @@ var LocalAIManager = class _LocalAIManager {
       throw error;
     }
   }
-  /**
-   * Generate embeddings - now with lazy initialization
-   */
-  async generateEmbedding(text) {
-    try {
-      await this.lazyInitEmbedding();
-      if (!this.embeddingModel) {
-        throw new Error("Failed to initialize embedding model");
-      }
-      logger9.info("Generating query embedding...");
-      const embedding = await this.embeddingModel.queryEmbed(text);
-      const dimensions = embedding.length;
-      logger9.info("Embedding generation complete", { dimensions });
-      return Array.from(embedding);
-    } catch (error) {
-      logger9.error("Embedding generation failed:", {
-        error: error instanceof Error ? error.message : String(error),
-        stack: error instanceof Error ? error.stack : void 0,
-        // Only access text.length if text exists
-        textLength: text?.length ?? "text is null"
-      });
-      throw error;
-    }
-  }
   /**
    * Describe image with lazy vision model initialization
    */
@@ -2923,26 +3003,6 @@ var LocalAIManager = class _LocalAIManager {
     }
     await this.mediumModelInitializingPromise;
   }
-  /**
-   * Lazy initialize the embedding model
-   */
-  async lazyInitEmbedding() {
-    if (this.embeddingInitialized) return;
-    if (!this.embeddingInitializingPromise) {
-      this.embeddingInitializingPromise = (async () => {
-        try {
-          await this.initializeEmbedding();
-          this.embeddingInitialized = true;
-          logger9.info("Embedding model initialized successfully");
-        } catch (error) {
-          logger9.error("Failed to initialize embedding model:", error);
-          this.embeddingInitializingPromise = null;
-          throw error;
-        }
-      })();
-    }
-    await this.embeddingInitializingPromise;
-  }
   /**
    * Lazy initialize the vision model
    */