feat(providers): add transformers

src/lib/models/providers/transformers/index.ts

@@ -0,0 +1,88 @@
+import { UIConfigField } from '@/lib/config/types';
+import { getConfiguredModelProviderById } from '@/lib/config/serverRegistry';
+import { Model, ModelList, ProviderMetadata } from '../../types';
+import BaseModelProvider from '../../base/provider';
+import BaseLLM from '../../base/llm';
+import BaseEmbedding from '../../base/embedding';
+import TransformerEmbedding from './transformerEmbedding';
+
+interface TransformersConfig {}
+
+const defaultEmbeddingModels: Model[] = [
+  {
+    name: 'all-MiniLM-L6-v2',
+    key: 'Xenova/all-MiniLM-L6-v2',
+  },
+  {
+    name: 'mxbai-embed-large-v1',
+    key: 'mixedbread-ai/mxbai-embed-large-v1',
+  },
+  {
+    name: 'nomic-embed-text-v1',
+    key: 'Xenova/nomic-embed-text-v1',
+  },
+];
+
+const providerConfigFields: UIConfigField[] = [];
+
+class TransformersProvider extends BaseModelProvider<TransformersConfig> {
+  constructor(id: string, name: string, config: TransformersConfig) {
+    super(id, name, config);
+  }
+
+  async getDefaultModels(): Promise<ModelList> {
+    return {
+      embedding: [...defaultEmbeddingModels],
+      chat: [],
+    };
+  }
+
+  async getModelList(): Promise<ModelList> {
+    const defaultModels = await this.getDefaultModels();
+    const configProvider = getConfiguredModelProviderById(this.id)!;
+
+    return {
+      embedding: [
+        ...defaultModels.embedding,
+        ...configProvider.embeddingModels,
+      ],
+      chat: [],
+    };
+  }
+
+  async loadChatModel(key: string): Promise<BaseLLM<any>> {
+    throw new Error('Transformers Provider does not support chat models.');
+  }
+
+  async loadEmbeddingModel(key: string): Promise<BaseEmbedding<any>> {
+    const modelList = await this.getModelList();
+    const exists = modelList.embedding.find((m) => m.key === key);
+
+    if (!exists) {
+      throw new Error(
+        'Error Loading OpenAI Embedding Model. Invalid Model Selected.',
+      );
+    }
+
+    return new TransformerEmbedding({
+      model: key,
+    });
+  }
+
+  static parseAndValidate(raw: any): TransformersConfig {
+    return {};
+  }
+
+  static getProviderConfigFields(): UIConfigField[] {
+    return providerConfigFields;
+  }
+
+  static getProviderMetadata(): ProviderMetadata {
+    return {
+      key: 'transformers',
+      name: 'Transformers',
+    };
+  }
+}
+
+export default TransformersProvider;

src/lib/models/providers/transformers/transformerEmbedding.ts

@@ -0,0 +1,43 @@
+import { Chunk } from '@/lib/types';
+import BaseEmbedding from '../../base/embedding';
+import { FeatureExtractionPipeline, pipeline } from '@huggingface/transformers';
+
+type TransformerConfig = {
+  model: string;
+};
+
+class TransformerEmbedding extends BaseEmbedding<TransformerConfig> {
+  private pipelinePromise: Promise<FeatureExtractionPipeline> | null = null;
+
+  constructor(protected config: TransformerConfig) {
+    super(config);
+  }
+
+  async embedText(texts: string[]): Promise<number[][]> {
+    return this.embed(texts);
+  }
+
+  async embedChunks(chunks: Chunk[]): Promise<number[][]> {
+    return this.embed(chunks.map((c) => c.content));
+  }
+
+  async embed(texts: string[]): Promise<number[][]> {
+    if (!this.pipelinePromise) {
+      this.pipelinePromise = (async () => {
+        const transformers = await import('@huggingface/transformers');
+        return (await transformers.pipeline(
+          'feature-extraction',
+          this.config.model,
+        )) as unknown as FeatureExtractionPipeline;
+      })();
+    }
+
+    const pipeline = await this.pipelinePromise;
+
+    const output = await pipeline(texts, { pooling: 'mean', normalize: true });
+
+    return output.tolist() as number[][];
+  }
+}
+
+export default TransformerEmbedding;