vecbox 0.2.2 → 0.2.3

package/src/providers/llamacpp.ts

@@ -1,4 +1,4 @@
-import { access, constants } from 'fs/promises';
+import { access, constants, readFile as fsReadFile } from 'fs/promises';
 import { join, resolve } from 'path';
 import { EmbeddingProvider } from '@providers/base/EmbeddingProvider';
 import type { EmbedConfig, EmbedInput, EmbedResult, BatchEmbedResult } from '@src/types/index';
@@ -13,8 +13,28 @@ import * as http from 'http';
 // Try to import native module
 let nativeModule: any = null;
 try {
-  nativeModule = require('../../native');
-  logger.info('Using native Llama.cpp module');
+  // Try different paths for native module
+  const possiblePaths = [
+    '../../native',  // Development
+    'vecbox/native',  // Installed as dependency
+    './native',      // Same directory
+    '../native',       // One level up
+    '../vecbox/native' // When installed via npm
+  ];
+  
+  for (const path of possiblePaths) {
+    try {
+      nativeModule = require(path);
+      logger.info(`Using native Llama.cpp module from: ${path}`);
+      break;
+    } catch (e) {
+      // Continue to next path
+    }
+  }
+  
+  if (!nativeModule) {
+    throw new Error('Native module not found in any path');
+  }
 } catch (error) {
   logger.warn('Native module not available, falling back to HTTP');
 }
@@ -88,25 +108,6 @@ export class LlamaCppProvider extends EmbeddingProvider {
     }
   }
 
-  private async loadGGUFModel(modelPath: string): Promise<Buffer> {
-    try {
-      logger.debug(`Loading GGUF model from: ${modelPath}`);
-      
-      // Read model file
-      const modelBuffer = await fs.readFile(modelPath);
-      
-      if (!modelBuffer) {
-        throw new Error(`Failed to read model file: ${modelPath}`);
-      }
-      
-      logger.debug(`Model file loaded, size: ${modelBuffer.length} bytes`);
-      return modelBuffer;
-    } catch (error) {
-      logger.error(`Failed to load GGUF model: ${error instanceof Error ? error.message : String(error)}`);
-      throw error;
-    }
-  }
-
   private generateEmbedding(modelBuffer: Buffer, text: string): number[] {
     // Use the loaded model to generate embedding
     logger.debug(`Generating embedding with model buffer (${modelBuffer.length} bytes)`);
@@ -142,8 +143,8 @@ export class LlamaCppProvider extends EmbeddingProvider {
         };
       }
 
-      // TODO: Implement direct Llama.cpp core usage in future
-      throw new Error('Direct Llama.cpp core integration not yet implemented. Please use HTTP fallback or wait for next version.');
+      // Fallback: return error if native module not available
+      throw new Error('Direct Llama.cpp integration requires native module. Please ensure native module is properly compiled.');
     } catch (error: unknown) {
       logger.error(`Llama.cpp embedding failed: ${(error instanceof Error ? error.message : String(error))}`);
       throw error;
@@ -178,42 +179,8 @@ export class LlamaCppProvider extends EmbeddingProvider {
         };
       }
 
-      // Fallback to HTTP batch processing
-      const texts = [];
-      for (const input of inputs) {
-        const text = await this.readInput(input);
-        if (text.trim()) {
-          texts.push(text);
-        }
-      }
-      
-      if (texts.length === 0) {
-        throw new Error('No valid texts to embed');
-      }
-
-      // For batch processing, use HTTP API
-      const modelPath = await this.getModelPath();
-      const requests = inputs.map((input, v) => ({
-        input: input.text || '',
-        model: modelPath,
-        pooling: 'mean',
-        normalize: 2
-      }));
-
-      // Execute batch requests (for now, do individual requests)
-      const embeddings: number[][] = [];
-      for (const request of requests) {
-        const result = await this.executeLlamaEmbedding([JSON.stringify(request)]);
-        const embedding = this.parseRawOutput(result.stdout);
-        embeddings.push(embedding);
-      }
-      
-      return {
-        embeddings,
-        dimensions: embeddings[0]?.length || 0,
-        model: this.getModel(),
-        provider: 'llamacpp',
-      };
+      // Fallback: return error if native module not available
+      throw new Error('Direct Llama.cpp integration requires native module. Please ensure native module is properly compiled.');
     } catch (error: unknown) {
       logger.error(`Llama.cpp batch embedding failed: ${(error instanceof Error ? error.message : String(error))}`);
       throw error;
@@ -237,4 +204,31 @@ export class LlamaCppProvider extends EmbeddingProvider {
   protected getModel(): string {
     return this.modelPath;
   }
+
+  private async getModelPath(): Promise<string> {
+    // If modelPath is already absolute, return as-is
+    if (this.modelPath.startsWith('/') || this.modelPath.startsWith('./')) {
+      return this.modelPath;
+    }
+    
+    // Try to resolve model path relative to current directory
+    const possiblePaths = [
+      resolve(this.modelPath),           // Current directory
+      join('core/models', this.modelPath),  // core/models subdirectory
+      join('models', this.modelPath),       // models subdirectory
+      join(__dirname, '../../core/models', this.modelPath), // Relative to dist
+    ];
+    
+    for (const path of possiblePaths) {
+      try {
+        await access(path, constants.F_OK);
+        return path;
+      } catch (e) {
+        // Continue to next path
+      }
+    }
+    
+    // Return original path if none found (will fail later with proper error)
+    return this.modelPath;
+  }
 }

package/src/providers/mistral.ts

@@ -8,6 +8,10 @@ const logger = Logger.createModuleLogger('mistral');
 export class MistralProvider extends EmbeddingProvider {
   private client: Mistral;
 
+  protected getModel(): string {
+    return this.config.model || 'mistral-embed';
+  }
+
   constructor(config: EmbedConfig) {
     super(config);