@wavequery/conductor/dist/esm/llm/huggingface.js

Modular LLM orchestration framework

export class HuggingFaceProvider {
    constructor(config) {
        this.pipeline = null;
        this.model = config.model;
    }
    async ensurePipeline() {
        if (!this.pipeline) {
            const transformers = await import("@xenova/transformers");
            this.pipeline = (await transformers.pipeline("text-generation", this.model));
        }
        return this.pipeline;
    }
    async complete(prompt, options = {}) {
        try {
            const pipe = await this.ensurePipeline();
            const response = await pipe(prompt, {
                max_new_tokens: options.maxTokens,
                temperature: options.temperature,
            });
            return {
                content: response[0].generated_text,
                usage: {
                    promptTokens: 0,
                    completionTokens: 0,
                    totalTokens: 0,
                },
                raw: response,
            };
        }
        catch (error) {
            throw new Error(`HuggingFace completion error: ${error.message}`);
        }
    }
    async completeWithFunctions(prompt, functions, options = {}) {
        // Similar to Anthropic, format functions as part of the prompt
        const functionsPrompt = `
      You must respond using one of these functions:
      ${JSON.stringify(functions, null, 2)}
      
      Your response must be a JSON object with a "name" field indicating the function
      and an "arguments" object containing the function parameters.
      
      Original request: ${prompt}
    `;
        try {
            const response = await this.complete(functionsPrompt, options);
            try {
                const parsedResponse = JSON.parse(response.content);
                return {
                    ...response,
                    functionCall: {
                        name: parsedResponse.name,
                        arguments: parsedResponse.arguments,
                    },
                };
            }
            catch (parseError) {
                throw new Error("Failed to parse function call response from HuggingFace model");
            }
        }
        catch (error) {
            throw new Error(`HuggingFace function completion error: ${error.message}`);
        }
    }
}
//# sourceMappingURL=huggingface.js.map