SmolLLM

LLM utility library

目前為 2025-03-04 提交的版本,檢視 最新版本

此腳本不應該直接安裝,它是一個供其他腳本使用的函式庫。欲使用本函式庫,請在腳本 metadata 寫上: // @require https://update.cn-greasyfork.org/scripts/528704/1546734/SmolLLM.js

您需要先安裝使用者腳本管理器擴展,如 TampermonkeyGreasemonkeyViolentmonkey 之後才能安裝該腳本。

You will need to install an extension such as Tampermonkey to install this script.

您需要先安裝使用者腳本管理器擴充功能,如 TampermonkeyViolentmonkey 後才能安裝該腳本。

您需要先安裝使用者腳本管理器擴充功能,如 TampermonkeyUserscripts 後才能安裝該腳本。

你需要先安裝一款使用者腳本管理器擴展,比如 Tampermonkey,才能安裝此腳本

您需要先安裝使用者腳本管理器擴充功能後才能安裝該腳本。

(我已經安裝了使用者腳本管理器,讓我安裝!)

你需要先安裝一款使用者樣式管理器擴展,比如 Stylus,才能安裝此樣式

你需要先安裝一款使用者樣式管理器擴展,比如 Stylus,才能安裝此樣式

你需要先安裝一款使用者樣式管理器擴展,比如 Stylus,才能安裝此樣式

你需要先安裝一款使用者樣式管理器擴展後才能安裝此樣式

你需要先安裝一款使用者樣式管理器擴展後才能安裝此樣式

你需要先安裝一款使用者樣式管理器擴展後才能安裝此樣式

(我已經安裝了使用者樣式管理器,讓我安裝!)

// ==UserScript==
// @name         SmolLLM
// @namespace    http://tampermonkey.net/
// @version      0.1.10
// @description  LLM utility library
// @author       RoCry
// @grant        GM_xmlhttpRequest
// @require https://update.greasyfork.org/scripts/528703/1546610/SimpleBalancer.js
// @license MIT
// ==/UserScript==

class SmolLLM {
  constructor() {
    // Ensure SimpleBalancer is available
    if (typeof SimpleBalancer === 'undefined') {
      throw new Error('SimpleBalancer is required for SmolLLM to work');
    }

    // Verify GM_xmlhttpRequest is available
    if (typeof GM_xmlhttpRequest === 'undefined') {
      throw new Error('GM_xmlhttpRequest is required for SmolLLM to work');
    }

    this.balancer = new SimpleBalancer();
    this.logger = console;

    // Buffer for incomplete SSE messages
    this.sseBuffer = '';

    // Track previously processed data length to avoid duplication
    this.processedLength = 0;
  }

  /**
   * Prepares request data based on the provider
   * 
   * @param {string} prompt - User prompt
   * @param {string} systemPrompt - System prompt 
   * @param {string} modelName - Model name
   * @param {string} providerName - Provider name (anthropic, openai, gemini)
   * @param {string} baseUrl - API base URL
   * @returns {Object} - {url, data} for the request
   */
  prepareRequestData(prompt, systemPrompt, modelName, providerName, baseUrl) {
    let url, data;

    if (providerName === 'anthropic') {
      url = `${baseUrl}/v1/messages`;
      data = {
        model: modelName,
        max_tokens: 4096,
        messages: [{ role: 'user', content: prompt }],
        stream: true
      };
      if (systemPrompt) {
        data.system = systemPrompt;
      }
    } else if (providerName === 'gemini') {
      url = `${baseUrl}/v1beta/models/${modelName}:streamGenerateContent?alt=sse`;
      data = {
        contents: [{ parts: [{ text: prompt }] }]
      };
      if (systemPrompt) {
        data.system_instruction = { parts: [{ text: systemPrompt }] };
      }
    } else {
      // OpenAI compatible APIs
      const messages = [];
      if (systemPrompt) {
        messages.push({ role: 'system', content: systemPrompt });
      }
      messages.push({ role: 'user', content: prompt });

      data = {
        messages: messages,
        model: modelName,
        stream: true
      };

      // Handle URL based on suffix
      if (baseUrl.endsWith('#')) {
        url = baseUrl.slice(0, -1); // Remove the # and use exact URL
      } else if (baseUrl.endsWith('/')) {
        url = `${baseUrl}chat/completions`; // Skip v1 prefix
      } else {
        url = `${baseUrl}/v1/chat/completions`; // Default pattern
      }
    }

    return { url, data };
  }

  /**
   * Prepares headers for authentication based on the provider
   * 
   * @param {string} providerName - Provider name
   * @param {string} apiKey - API key
   * @returns {Object} - Request headers
   */
  prepareHeaders(providerName, apiKey) {
    const headers = {
      'Content-Type': 'application/json'
    };

    if (providerName === 'anthropic') {
      headers['X-API-Key'] = apiKey;
      headers['Anthropic-Version'] = '2023-06-01';
    } else if (providerName === 'gemini') {
      headers['X-Goog-Api-Key'] = apiKey;
    } else {
      headers['Authorization'] = `Bearer ${apiKey}`;
    }

    return headers;
  }

  /**
   * Process SSE stream data for different providers
   * 
   * @param {string} chunk - Data chunk from SSE
   * @param {string} providerName - Provider name
   * @returns {string|null} - Extracted text content or null
   */
  processStreamChunk(chunk, providerName) {
    if (!chunk || chunk === '[DONE]') return null;

    try {
      console.log(`Processing chunk for ${providerName}:`, chunk.substring(0, 100) + (chunk.length > 100 ? '...' : ''));
      const data = JSON.parse(chunk);

      if (providerName === 'gemini') {
        const candidates = data.candidates || [];
        if (candidates.length > 0 && candidates[0].content) {
          const parts = candidates[0].content.parts;
          if (parts && parts.length > 0) {
            return parts[0].text || '';
          }
        }
        return null;
      } else if (providerName === 'anthropic') {
        // Handle content_block_delta which contains the actual text
        if (data.type === 'content_block_delta') {
          const delta = data.delta || {};
          if (delta.type === 'text_delta' || delta.text) {
            return delta.text || '';
          }
        }
        // Anthropic sends various event types - only some contain text
        return null;
      } else {
        // OpenAI compatible format
        const choice = (data.choices || [{}])[0];
        if (choice.finish_reason !== null && choice.finish_reason !== undefined) {
          return null; // End of generation
        }
        return choice.delta && choice.delta.content ? choice.delta.content : null;
      }
    } catch (e) {
      console.error(`Error parsing chunk: ${e.message}, chunk: ${chunk}`);
      return null;
    }
  }

  /**
   * Process SSE data and extract complete messages
   * 
   * @param {string} rawData - Raw data received from SSE
   * @param {string} providerName - Provider name for chunk processing
   * @param {Function} handleText - Callback for processed text chunks
   */
  processSseData(rawData, providerName, handleText) {
    // Append new data to existing buffer
    this.sseBuffer += rawData;

    // Find complete SSE messages (data: ... followed by two newlines)
    const regex = /data: (.*?)(?:\n\n|\r\n\r\n)/g;
    let match;
    let newBuffer = this.sseBuffer;

    while ((match = regex.exec(this.sseBuffer)) !== null) {
      try {
        const content = match[1];
        const textChunk = this.processStreamChunk(content, providerName);

        if (textChunk) {
          handleText(textChunk);
        }

        // Remove processed message from buffer
        newBuffer = newBuffer.substring(match.index + match[0].length);
      } catch (e) {
        console.error('Error processing SSE message:', e);
      }
    }

    // Save the remaining unprocessed data for next time
    this.sseBuffer = newBuffer;
  }

  /**
   * Makes a request to the LLM API and handles streaming responses
   * 
   * @param {Object} params - Request parameters
   * @returns {Promise<string>} - Full response text
   */
  async askLLM({
    prompt,
    providerName,
    systemPrompt = '',
    model,
    apiKey,
    baseUrl,
    handler = null,
    timeout = 60000
  }) {
    if (!prompt || !providerName || !model || !apiKey || !baseUrl) {
      throw new Error('Required parameters missing');
    }

    // Use balancer to choose API key and base URL pair
    [apiKey, baseUrl] = this.balancer.choosePair(apiKey, baseUrl);

    const { url, data } = this.prepareRequestData(
      prompt, systemPrompt, model, providerName, baseUrl
    );

    const headers = this.prepareHeaders(providerName, apiKey);

    // Log request info (with masked API key)
    const apiKeyPreview = `${apiKey.slice(0, 5)}...${apiKey.slice(-4)}`;
    this.logger.info(
      `Sending request to ${url} with model=${model}, api_key=${apiKeyPreview}, prompt_length=${prompt.length}`
    );

    // Additional debug info
    this.logger.debug(`Provider: ${providerName}, Request data:`, JSON.stringify(data).substring(0, 500));

    // Reset SSE buffer and processed length before new request
    this.sseBuffer = '';
    this.processedLength = 0;

    return new Promise((resolve, reject) => {
      let responseText = '';
      let timeoutId;

      // Set timeout
      if (timeout) {
        timeoutId = setTimeout(() => {
          reject(new Error(`Request timed out after ${timeout}ms`));
        }, timeout);
      }

      GM_xmlhttpRequest({
        method: 'POST',
        url: url,
        headers: headers,
        data: JSON.stringify(data),
        responseType: 'text', // Use 'text' instead of unsupported 'stream'
        onload: (response) => {
          if (response.status !== 200) {
            clearTimeout(timeoutId);
            reject(new Error(`API request failed: ${response.status} - ${response.responseText}`));
          } else {
            // In case the response isn't streamed, process it all at once
            if (responseText.length === 0 && response.responseText) {
              this.processSseData(response.responseText, providerName, (textChunk) => {
                responseText += textChunk;
                if (handler && typeof handler === 'function') {
                  handler(textChunk);
                }
              });
            }
            clearTimeout(timeoutId);
            resolve(responseText);
          }
        },
        onreadystatechange: (state) => {
          if (state.readyState === 4) {
            // Request completed
            clearTimeout(timeoutId);
            resolve(responseText);
          }
        },
        onprogress: (response) => {
          // Get only the new data since last progress event
          const newData = response.responseText.substring(this.processedLength);
          this.processedLength = response.responseText.length;

          if (!newData) return;

          // Process only the new data with our improved SSE parser
          this.processSseData(newData, providerName, (textChunk) => {
            responseText += textChunk;
            if (handler && typeof handler === 'function') {
              handler(textChunk);
            }
          });
        },
        onerror: (error) => {
          clearTimeout(timeoutId);
          console.error('Request error:', error);
          reject(new Error(`Request failed: ${error.error || JSON.stringify(error)}`));
        },
        ontimeout: () => {
          clearTimeout(timeoutId);
          reject(new Error(`Request timed out after ${timeout}ms`));
        }
      });
    });
  }
}

// Make it available globally
window.SmolLLM = SmolLLM;

// Export for module systems if needed
if (typeof module !== 'undefined') {
  module.exports = SmolLLM;
}