/**
* Sentence Transformers Embeddings Provider
*
* Uses @xenova/transformers to provide local embeddings without requiring API keys.
* This is perfect for self-hosted MCP servers and eliminates external dependencies.
*/
import log from '../../logger.js';
/**
* LangChain-compatible embeddings class using sentence-transformers
*/
export class SentenceTransformersEmbeddings {
private model: any = null;
private modelName: string;
private isInitialized: boolean = false;
private transformers: any = null;
constructor(options: { modelName?: string } = {}) {
// Use a lightweight, fast model by default
this.modelName = options.modelName || 'Xenova/all-MiniLM-L6-v2';
}
/**
* Initialize the transformers library dynamically
*/
private async initializeTransformers(): Promise<void> {
if (this.transformers) {
return;
}
try {
// Use eval to avoid CommonJS/ESM conflict during compilation
const importTransformers = new Function(
'return import("@xenova/transformers")'
);
this.transformers = await importTransformers();
} catch (error) {
log.error('Error importing @xenova/transformers:', error);
throw new Error(
`Failed to import @xenova/transformers: ${error instanceof Error ? error.message : String(error)}`
);
}
}
/**
* Initialize the model lazily
*/
private async initializeModel(): Promise<void> {
if (this.isInitialized && this.model) {
return;
}
await this.initializeTransformers();
log.info(`Initializing sentence-transformers model: ${this.modelName}`);
try {
this.model = await this.transformers.pipeline(
'feature-extraction',
this.modelName
);
this.isInitialized = true;
log.info(
`Successfully initialized sentence-transformers model: ${this.modelName}`
);
} catch (error) {
log.error('Error initializing sentence-transformers model:', error);
throw new Error(
`Failed to initialize sentence-transformers model: ${error instanceof Error ? error.message : String(error)}`
);
}
}
/**
* Generate embeddings for a single text (LangChain interface)
*/
async embedQuery(text: string): Promise<number[]> {
await this.initializeModel();
if (!this.model) {
throw new Error('Model not initialized');
}
try {
const result = await this.model(text, {
pooling: 'mean',
normalize: true,
});
// Convert tensor to array
const embeddings = Array.from(result.data) as number[];
return embeddings;
} catch (error) {
log.error('Error generating embeddings:', error);
throw new Error(
`Failed to generate embeddings: ${error instanceof Error ? error.message : String(error)}`
);
}
}
/**
* Generate embeddings for multiple texts (LangChain interface)
*/
async embedDocuments(texts: string[]): Promise<number[][]> {
await this.initializeModel();
if (!this.model) {
throw new Error('Model not initialized');
}
try {
const embeddings: number[][] = [];
// Process texts in batches to avoid memory issues
const batchSize = 10;
for (let i = 0; i < texts.length; i += batchSize) {
const batch = texts.slice(i, i + batchSize);
for (const text of batch) {
const result = await this.model(text, {
pooling: 'mean',
normalize: true,
});
const embedding = Array.from(result.data) as number[];
embeddings.push(embedding);
}
// Log progress for large batches
if (texts.length > batchSize) {
log.info(
`Processed ${Math.min(i + batchSize, texts.length)}/${texts.length} documents`
);
}
}
return embeddings;
} catch (error) {
log.error('Error generating document embeddings:', error);
throw new Error(
`Failed to generate document embeddings: ${error instanceof Error ? error.message : String(error)}`
);
}
}
}
