onbuzz 4.9.13 → 4.10.0

package/src/services/ollamaService.js

@@ -1,452 +1,452 @@
-/**
- * @file ollamaService.js
- * @description Ollama integration service for local/offline LLM inference.
- * Wraps the ollama-js library to provide chat, streaming, and model discovery
- * capabilities that integrate with the existing AI service architecture.
- *
- * Ollama models are free to use (no billing) and run entirely offline.
- */
-
-import { Ollama } from 'ollama';
-
-const DEFAULT_OLLAMA_HOST = 'http://127.0.0.1:11434';
-const MODEL_CACHE_TTL = 60_000; // 1 minute cache for model list
-const HEALTH_CHECK_TIMEOUT = 3000; // 3 second timeout for health check
-const OLLAMA_MODEL_PREFIX = 'ollama-';
-
-/**
- * OllamaService - Local LLM inference via Ollama
- */
-class OllamaService {
-  constructor(config = {}, logger = null) {
-    this.logger = logger;
-    this.host = config.ollamaHost || DEFAULT_OLLAMA_HOST;
-    this.client = new Ollama({ host: this.host });
-    this.enabled = config.ollamaEnabled !== false; // Enabled by default
-
-    // Model cache
-    this._modelCache = null;
-    this._modelCacheTime = 0;
-
-    // Connection state
-    this._isAvailable = null; // null = unknown
-    this._lastHealthCheck = 0;
-  }
-
-  /**
-   * Update Ollama host URL
-   */
-  setHost(host) {
-    this.host = host || DEFAULT_OLLAMA_HOST;
-    this.client = new Ollama({ host: this.host });
-    this._isAvailable = null;
-    this._modelCache = null;
-    this.logger?.info(`[Ollama] Host updated to ${this.host}`);
-  }
-
-  /**
-   * Check if Ollama server is reachable
-   */
-  async isAvailable() {
-    // Cache health check for 10 seconds
-    if (this._isAvailable !== null && Date.now() - this._lastHealthCheck < 10_000) {
-      return this._isAvailable;
-    }
-
-    try {
-      const controller = new AbortController();
-      const timeoutId = setTimeout(() => controller.abort(), HEALTH_CHECK_TIMEOUT);
-      const response = await fetch(`${this.host}/api/version`, {
-        signal: controller.signal
-      });
-      clearTimeout(timeoutId);
-      this._isAvailable = response.ok;
-    } catch {
-      this._isAvailable = false;
-    }
-
-    this._lastHealthCheck = Date.now();
-    return this._isAvailable;
-  }
-
-  /**
-   * List available local models from Ollama
-   * @returns {Array<Object>} Model list with normalized specs
-   */
-  async listModels() {
-    // Return cached if fresh
-    if (this._modelCache && Date.now() - this._modelCacheTime < MODEL_CACHE_TTL) {
-      return this._modelCache;
-    }
-
-    if (!this.enabled) return [];
-
-    try {
-      const available = await this.isAvailable();
-      if (!available) return [];
-
-      const response = await this.client.list();
-      const models = (response.models || []).map(m => this._normalizeModel(m));
-
-      this._modelCache = models;
-      this._modelCacheTime = Date.now();
-
-      this.logger?.info(`[Ollama] Discovered ${models.length} local models`);
-      return models;
-    } catch (err) {
-      this.logger?.warn(`[Ollama] Failed to list models: ${err.message}`);
-      return this._modelCache || [];
-    }
-  }
-
-  /**
-   * Normalize Ollama model info to match platform model schema
-   * @private
-   */
-  _normalizeModel(ollamaModel) {
-    const name = ollamaModel.name; // e.g. "llama3.1:8b", "codellama:13b"
-    const modelId = OLLAMA_MODEL_PREFIX + name.replace(/[:/]/g, '-'); // ollama-llama3.1-8b
-    const details = ollamaModel.details || {};
-    const sizeGB = ollamaModel.size ? (ollamaModel.size / 1e9).toFixed(1) : '?';
-
-    // Estimate context window from parameter size and quantization
-    const paramSize = details.parameter_size || '';
-    const contextWindow = this._estimateContextWindow(paramSize, name);
-
-    return {
-      name: modelId,
-      ollamaName: name, // Original name for API calls
-      displayName: `${name} (${sizeGB}GB)`,
-      provider: 'ollama',
-      type: 'chat',
-      api_type: ['chat'],
-      contextWindow,
-      maxTokens: Math.min(4096, Math.floor(contextWindow * 0.25)),
-      pricing: { input: 0, output: 0 }, // Free - local inference
-      deprecated: false,
-      available: true,
-      local: true,
-      offline: true,
-      details: {
-        family: details.family || null,
-        parameterSize: paramSize,
-        quantization: details.quantization_level || null,
-        format: details.format || null,
-        sizeBytes: ollamaModel.size || 0
-      }
-    };
-  }
-
-  /**
-   * Estimate context window based on model info
-   * @private
-   */
-  _estimateContextWindow(paramSize, name) {
-    // Known models with specific context windows
-    const nameLower = name.toLowerCase();
-    if (nameLower.includes('llama3') || nameLower.includes('llama-3')) return 128000;
-    if (nameLower.includes('llama2') || nameLower.includes('llama-2')) return 4096;
-    if (nameLower.includes('mistral')) return 32768;
-    if (nameLower.includes('mixtral')) return 32768;
-    if (nameLower.includes('phi')) return 16384;
-    if (nameLower.includes('gemma2') || nameLower.includes('gemma-2')) return 8192;
-    if (nameLower.includes('qwen')) return 32768;
-    if (nameLower.includes('deepseek')) return 128000;
-    if (nameLower.includes('codellama')) return 16384;
-    if (nameLower.includes('command-r')) return 128000;
-
-    // Default based on parameter size
-    const sizeMatch = paramSize.match(/([\d.]+)/);
-    if (sizeMatch) {
-      const params = parseFloat(sizeMatch[1]);
-      if (params >= 70) return 8192;
-      if (params >= 13) return 8192;
-      return 4096;
-    }
-
-    return 4096; // Conservative default
-  }
-
-  /**
-   * Send a chat completion request to Ollama
-   * @param {string} modelId - Normalized model ID (ollama-xxx)
-   * @param {Array} messages - Conversation messages [{role, content}]
-   * @param {Object} options - Options (temperature, maxTokens, systemPrompt)
-   * @returns {Object} Response matching AIService format
-   */
-  async sendMessage(modelId, messages, options = {}) {
-    const ollamaName = this._resolveModelName(modelId);
-
-    // Build message array with system prompt
-    const ollamaMessages = [];
-    if (options.systemPrompt) {
-      ollamaMessages.push({ role: 'system', content: options.systemPrompt });
-    }
-
-    for (const msg of messages) {
-      ollamaMessages.push({
-        role: msg.role === 'assistant' ? 'assistant' : (msg.role === 'system' ? 'system' : 'user'),
-        content: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content)
-      });
-    }
-
-    const startTime = Date.now();
-
-    try {
-      const response = await this.client.chat({
-        model: ollamaName,
-        messages: ollamaMessages,
-        stream: false,
-        options: {
-          temperature: options.temperature ?? 0.7,
-          num_predict: options.maxTokens || 4096,
-        }
-      });
-
-      const duration = Date.now() - startTime;
-
-      // Estimate token counts from Ollama's eval metrics
-      const promptTokens = response.prompt_eval_count || this._estimateTokens(ollamaMessages);
-      const completionTokens = response.eval_count || this._estimateTokens([{ content: response.message?.content || '' }]);
-
-      this.logger?.info(`[Ollama] ${ollamaName} responded in ${duration}ms (${promptTokens}+${completionTokens} tokens)`);
-
-      return {
-        choices: [{
-          message: { content: response.message?.content || '' },
-          finish_reason: response.done ? 'stop' : 'length'
-        }],
-        model: modelId,
-        usage: {
-          prompt_tokens: promptTokens,
-          completion_tokens: completionTokens,
-          total_tokens: promptTokens + completionTokens
-        }
-      };
-    } catch (err) {
-      this.logger?.error(`[Ollama] Chat error for ${ollamaName}: ${err.message}`);
-      throw new Error(`Ollama error (${ollamaName}): ${err.message}`);
-    }
-  }
-
-  /**
-   * Send a streaming chat request to Ollama
-   * @param {string} modelId - Normalized model ID
-   * @param {Array} messages - Conversation messages
-   * @param {Object} options - Options including onChunk, onDone, onError callbacks
-   * @returns {Object} Final response with content and metadata
-   */
-  async sendMessageStream(modelId, messages, options = {}) {
-    const ollamaName = this._resolveModelName(modelId);
-    const { onChunk, onDone, onError } = options;
-
-    // Build message array
-    const ollamaMessages = [];
-    if (options.systemPrompt) {
-      ollamaMessages.push({ role: 'system', content: options.systemPrompt });
-    }
-
-    for (const msg of messages) {
-      ollamaMessages.push({
-        role: msg.role === 'assistant' ? 'assistant' : (msg.role === 'system' ? 'system' : 'user'),
-        content: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content)
-      });
-    }
-
-    const startTime = Date.now();
-    let fullContent = '';
-    let promptTokens = 0;
-    let completionTokens = 0;
-
-    try {
-      const stream = await this.client.chat({
-        model: ollamaName,
-        messages: ollamaMessages,
-        stream: true,
-        options: {
-          temperature: options.temperature ?? 0.7,
-          num_predict: options.maxTokens || 4096,
-        }
-      });
-
-      if (!stream) {
-        throw new Error(`Ollama returned no stream for model "${ollamaName}". The model may not be loaded or available.`);
-      }
-
-      for await (const chunk of stream) {
-        const content = chunk.message?.content || '';
-        if (content) {
-          fullContent += content;
-          if (onChunk) {
-            onChunk({ content, type: 'chunk' });
-          }
-        }
-
-        // Final chunk has eval counts
-        if (chunk.done) {
-          promptTokens = chunk.prompt_eval_count || this._estimateTokens(ollamaMessages);
-          completionTokens = chunk.eval_count || this._estimateTokens([{ content: fullContent }]);
-        }
-      }
-
-      const duration = Date.now() - startTime;
-      const usage = {
-        prompt_tokens: promptTokens,
-        completion_tokens: completionTokens,
-        total_tokens: promptTokens + completionTokens
-      };
-
-      this.logger?.info(`[Ollama] ${ollamaName} streamed in ${duration}ms (${usage.total_tokens} tokens)`);
-
-      const result = {
-        content: fullContent,
-        model: modelId,
-        tokenUsage: usage,
-        finishReason: 'stop'
-      };
-
-      if (onDone) {
-        onDone(result);
-      }
-
-      return result;
-    } catch (err) {
-      this.logger?.error(`[Ollama] Stream error for ${ollamaName}: ${err.message}`);
-      if (onError) {
-        onError(err);
-      }
-      throw new Error(`Ollama streaming error (${ollamaName}): ${err.message}`);
-    }
-  }
-
-  /**
-   * Resolve normalized model ID back to Ollama model name
-   * @private
-   */
-  _resolveModelName(modelId) {
-    // If it has the prefix, strip it and restore colons
-    if (modelId.startsWith(OLLAMA_MODEL_PREFIX)) {
-      const stripped = modelId.slice(OLLAMA_MODEL_PREFIX.length);
-      // Restore the last dash to colon for tag separator (e.g., llama3.1-8b → llama3.1:8b)
-      // Use cached model list for accurate mapping
-      if (this._modelCache) {
-        const cached = this._modelCache.find(m => m.name === modelId);
-        if (cached) return cached.ollamaName;
-      }
-      // Fallback: restore last dash before a version-like segment to colon
-      return stripped.replace(/-(\d+[bgBG]?)$/, ':$1')
-                     .replace(/-latest$/, ':latest');
-    }
-    return modelId;
-  }
-
-  /**
-   * Estimate token count from messages (fallback when Ollama doesn't report)
-   * @private
-   */
-  _estimateTokens(messages) {
-    let chars = 0;
-    for (const msg of messages) {
-      const content = msg.content || '';
-      chars += typeof content === 'string' ? content.length : JSON.stringify(content).length;
-    }
-    return Math.ceil(chars / 3.5); // ~3.5 chars per token average
-  }
-
-  /**
-   * Check if a model ID is an Ollama model
-   * @param {string} modelId - Model identifier
-   * @returns {boolean}
-   */
-  static isOllamaModel(modelId) {
-    return typeof modelId === 'string' && modelId.startsWith(OLLAMA_MODEL_PREFIX);
-  }
-
-  /**
-   * Pull a model from Ollama registry
-   * @param {string} modelName - Model to pull (e.g., "llama3.1:8b")
-   * @param {Function} onProgress - Progress callback
-   */
-  async pullModel(modelName, onProgress = null) {
-    this.logger?.info(`[Ollama] Pulling model: ${modelName}`);
-
-    try {
-      const stream = await this.client.pull({ model: modelName, stream: true });
-
-      if (!stream) {
-        throw new Error(`Ollama returned no stream when pulling "${modelName}". Check that Ollama is running.`);
-      }
-
-      for await (const progress of stream) {
-        if (onProgress) {
-          onProgress({
-            status: progress.status,
-            total: progress.total,
-            completed: progress.completed,
-            percent: progress.total ? Math.round((progress.completed / progress.total) * 100) : null
-          });
-        }
-      }
-
-      // Invalidate model cache
-      this._modelCache = null;
-      this.logger?.info(`[Ollama] Successfully pulled ${modelName}`);
-      return { success: true, model: modelName };
-    } catch (err) {
-      this.logger?.error(`[Ollama] Pull failed for ${modelName}: ${err.message}`);
-      throw err;
-    }
-  }
-
-  /**
-   * Delete a model from Ollama
-   * @param {string} modelName - Model to delete
-   */
-  async deleteModel(modelName) {
-    try {
-      await this.client.delete({ model: modelName });
-      this._modelCache = null;
-      this.logger?.info(`[Ollama] Deleted model: ${modelName}`);
-      return { success: true };
-    } catch (err) {
-      this.logger?.error(`[Ollama] Delete failed for ${modelName}: ${err.message}`);
-      throw err;
-    }
-  }
-
-  /**
-   * Get model info from Ollama
-   * @param {string} modelName - Model name
-   */
-  async getModelInfo(modelName) {
-    try {
-      const info = await this.client.show({ model: modelName });
-      return {
-        success: true,
-        info: {
-          name: modelName,
-          license: info.license,
-          modelfile: info.modelfile,
-          parameters: info.parameters,
-          template: info.template,
-          details: info.details
-        }
-      };
-    } catch (err) {
-      return { success: false, error: err.message };
-    }
-  }
-}
-
-// Singleton
-let instance = null;
-
-export function getOllamaService(config = {}, logger = null) {
-  if (!instance) {
-    instance = new OllamaService(config, logger);
-  }
-  return instance;
-}
-
-export { OllamaService, OLLAMA_MODEL_PREFIX };
-export default OllamaService;
+/**
+ * @file ollamaService.js
+ * @description Ollama integration service for local/offline LLM inference.
+ * Wraps the ollama-js library to provide chat, streaming, and model discovery
+ * capabilities that integrate with the existing AI service architecture.
+ *
+ * Ollama models are free to use (no billing) and run entirely offline.
+ */
+
+import { Ollama } from 'ollama';
+
+const DEFAULT_OLLAMA_HOST = 'http://127.0.0.1:11434';
+const MODEL_CACHE_TTL = 60_000; // 1 minute cache for model list
+const HEALTH_CHECK_TIMEOUT = 3000; // 3 second timeout for health check
+const OLLAMA_MODEL_PREFIX = 'ollama-';
+
+/**
+ * OllamaService - Local LLM inference via Ollama
+ */
+class OllamaService {
+  constructor(config = {}, logger = null) {
+    this.logger = logger;
+    this.host = config.ollamaHost || DEFAULT_OLLAMA_HOST;
+    this.client = new Ollama({ host: this.host });
+    this.enabled = config.ollamaEnabled !== false; // Enabled by default
+
+    // Model cache
+    this._modelCache = null;
+    this._modelCacheTime = 0;
+
+    // Connection state
+    this._isAvailable = null; // null = unknown
+    this._lastHealthCheck = 0;
+  }
+
+  /**
+   * Update Ollama host URL
+   */
+  setHost(host) {
+    this.host = host || DEFAULT_OLLAMA_HOST;
+    this.client = new Ollama({ host: this.host });
+    this._isAvailable = null;
+    this._modelCache = null;
+    this.logger?.info(`[Ollama] Host updated to ${this.host}`);
+  }
+
+  /**
+   * Check if Ollama server is reachable
+   */
+  async isAvailable() {
+    // Cache health check for 10 seconds
+    if (this._isAvailable !== null && Date.now() - this._lastHealthCheck < 10_000) {
+      return this._isAvailable;
+    }
+
+    try {
+      const controller = new AbortController();
+      const timeoutId = setTimeout(() => controller.abort(), HEALTH_CHECK_TIMEOUT);
+      const response = await fetch(`${this.host}/api/version`, {
+        signal: controller.signal
+      });
+      clearTimeout(timeoutId);
+      this._isAvailable = response.ok;
+    } catch {
+      this._isAvailable = false;
+    }
+
+    this._lastHealthCheck = Date.now();
+    return this._isAvailable;
+  }
+
+  /**
+   * List available local models from Ollama
+   * @returns {Array<Object>} Model list with normalized specs
+   */
+  async listModels() {
+    // Return cached if fresh
+    if (this._modelCache && Date.now() - this._modelCacheTime < MODEL_CACHE_TTL) {
+      return this._modelCache;
+    }
+
+    if (!this.enabled) return [];
+
+    try {
+      const available = await this.isAvailable();
+      if (!available) return [];
+
+      const response = await this.client.list();
+      const models = (response.models || []).map(m => this._normalizeModel(m));
+
+      this._modelCache = models;
+      this._modelCacheTime = Date.now();
+
+      this.logger?.info(`[Ollama] Discovered ${models.length} local models`);
+      return models;
+    } catch (err) {
+      this.logger?.warn(`[Ollama] Failed to list models: ${err.message}`);
+      return this._modelCache || [];
+    }
+  }
+
+  /**
+   * Normalize Ollama model info to match platform model schema
+   * @private
+   */
+  _normalizeModel(ollamaModel) {
+    const name = ollamaModel.name; // e.g. "llama3.1:8b", "codellama:13b"
+    const modelId = OLLAMA_MODEL_PREFIX + name.replace(/[:/]/g, '-'); // ollama-llama3.1-8b
+    const details = ollamaModel.details || {};
+    const sizeGB = ollamaModel.size ? (ollamaModel.size / 1e9).toFixed(1) : '?';
+
+    // Estimate context window from parameter size and quantization
+    const paramSize = details.parameter_size || '';
+    const contextWindow = this._estimateContextWindow(paramSize, name);
+
+    return {
+      name: modelId,
+      ollamaName: name, // Original name for API calls
+      displayName: `${name} (${sizeGB}GB)`,
+      provider: 'ollama',
+      type: 'chat',
+      api_type: ['chat'],
+      contextWindow,
+      maxTokens: Math.min(4096, Math.floor(contextWindow * 0.25)),
+      pricing: { input: 0, output: 0 }, // Free - local inference
+      deprecated: false,
+      available: true,
+      local: true,
+      offline: true,
+      details: {
+        family: details.family || null,
+        parameterSize: paramSize,
+        quantization: details.quantization_level || null,
+        format: details.format || null,
+        sizeBytes: ollamaModel.size || 0
+      }
+    };
+  }
+
+  /**
+   * Estimate context window based on model info
+   * @private
+   */
+  _estimateContextWindow(paramSize, name) {
+    // Known models with specific context windows
+    const nameLower = name.toLowerCase();
+    if (nameLower.includes('llama3') || nameLower.includes('llama-3')) return 128000;
+    if (nameLower.includes('llama2') || nameLower.includes('llama-2')) return 4096;
+    if (nameLower.includes('mistral')) return 32768;
+    if (nameLower.includes('mixtral')) return 32768;
+    if (nameLower.includes('phi')) return 16384;
+    if (nameLower.includes('gemma2') || nameLower.includes('gemma-2')) return 8192;
+    if (nameLower.includes('qwen')) return 32768;
+    if (nameLower.includes('deepseek')) return 128000;
+    if (nameLower.includes('codellama')) return 16384;
+    if (nameLower.includes('command-r')) return 128000;
+
+    // Default based on parameter size
+    const sizeMatch = paramSize.match(/([\d.]+)/);
+    if (sizeMatch) {
+      const params = parseFloat(sizeMatch[1]);
+      if (params >= 70) return 8192;
+      if (params >= 13) return 8192;
+      return 4096;
+    }
+
+    return 4096; // Conservative default
+  }
+
+  /**
+   * Send a chat completion request to Ollama
+   * @param {string} modelId - Normalized model ID (ollama-xxx)
+   * @param {Array} messages - Conversation messages [{role, content}]
+   * @param {Object} options - Options (temperature, maxTokens, systemPrompt)
+   * @returns {Object} Response matching AIService format
+   */
+  async sendMessage(modelId, messages, options = {}) {
+    const ollamaName = this._resolveModelName(modelId);
+
+    // Build message array with system prompt
+    const ollamaMessages = [];
+    if (options.systemPrompt) {
+      ollamaMessages.push({ role: 'system', content: options.systemPrompt });
+    }
+
+    for (const msg of messages) {
+      ollamaMessages.push({
+        role: msg.role === 'assistant' ? 'assistant' : (msg.role === 'system' ? 'system' : 'user'),
+        content: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content)
+      });
+    }
+
+    const startTime = Date.now();
+
+    try {
+      const response = await this.client.chat({
+        model: ollamaName,
+        messages: ollamaMessages,
+        stream: false,
+        options: {
+          temperature: options.temperature ?? 0.7,
+          num_predict: options.maxTokens || 4096,
+        }
+      });
+
+      const duration = Date.now() - startTime;
+
+      // Estimate token counts from Ollama's eval metrics
+      const promptTokens = response.prompt_eval_count || this._estimateTokens(ollamaMessages);
+      const completionTokens = response.eval_count || this._estimateTokens([{ content: response.message?.content || '' }]);
+
+      this.logger?.info(`[Ollama] ${ollamaName} responded in ${duration}ms (${promptTokens}+${completionTokens} tokens)`);
+
+      return {
+        choices: [{
+          message: { content: response.message?.content || '' },
+          finish_reason: response.done ? 'stop' : 'length'
+        }],
+        model: modelId,
+        usage: {
+          prompt_tokens: promptTokens,
+          completion_tokens: completionTokens,
+          total_tokens: promptTokens + completionTokens
+        }
+      };
+    } catch (err) {
+      this.logger?.error(`[Ollama] Chat error for ${ollamaName}: ${err.message}`);
+      throw new Error(`Ollama error (${ollamaName}): ${err.message}`, { cause: err });
+    }
+  }
+
+  /**
+   * Send a streaming chat request to Ollama
+   * @param {string} modelId - Normalized model ID
+   * @param {Array} messages - Conversation messages
+   * @param {Object} options - Options including onChunk, onDone, onError callbacks
+   * @returns {Object} Final response with content and metadata
+   */
+  async sendMessageStream(modelId, messages, options = {}) {
+    const ollamaName = this._resolveModelName(modelId);
+    const { onChunk, onDone, onError } = options;
+
+    // Build message array
+    const ollamaMessages = [];
+    if (options.systemPrompt) {
+      ollamaMessages.push({ role: 'system', content: options.systemPrompt });
+    }
+
+    for (const msg of messages) {
+      ollamaMessages.push({
+        role: msg.role === 'assistant' ? 'assistant' : (msg.role === 'system' ? 'system' : 'user'),
+        content: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content)
+      });
+    }
+
+    const startTime = Date.now();
+    let fullContent = '';
+    let promptTokens = 0;
+    let completionTokens = 0;
+
+    try {
+      const stream = await this.client.chat({
+        model: ollamaName,
+        messages: ollamaMessages,
+        stream: true,
+        options: {
+          temperature: options.temperature ?? 0.7,
+          num_predict: options.maxTokens || 4096,
+        }
+      });
+
+      if (!stream) {
+        throw new Error(`Ollama returned no stream for model "${ollamaName}". The model may not be loaded or available.`);
+      }
+
+      for await (const chunk of stream) {
+        const content = chunk.message?.content || '';
+        if (content) {
+          fullContent += content;
+          if (onChunk) {
+            onChunk({ content, type: 'chunk' });
+          }
+        }
+
+        // Final chunk has eval counts
+        if (chunk.done) {
+          promptTokens = chunk.prompt_eval_count || this._estimateTokens(ollamaMessages);
+          completionTokens = chunk.eval_count || this._estimateTokens([{ content: fullContent }]);
+        }
+      }
+
+      const duration = Date.now() - startTime;
+      const usage = {
+        prompt_tokens: promptTokens,
+        completion_tokens: completionTokens,
+        total_tokens: promptTokens + completionTokens
+      };
+
+      this.logger?.info(`[Ollama] ${ollamaName} streamed in ${duration}ms (${usage.total_tokens} tokens)`);
+
+      const result = {
+        content: fullContent,
+        model: modelId,
+        tokenUsage: usage,
+        finishReason: 'stop'
+      };
+
+      if (onDone) {
+        onDone(result);
+      }
+
+      return result;
+    } catch (err) {
+      this.logger?.error(`[Ollama] Stream error for ${ollamaName}: ${err.message}`);
+      if (onError) {
+        onError(err);
+      }
+      throw new Error(`Ollama streaming error (${ollamaName}): ${err.message}`, { cause: err });
+    }
+  }
+
+  /**
+   * Resolve normalized model ID back to Ollama model name
+   * @private
+   */
+  _resolveModelName(modelId) {
+    // If it has the prefix, strip it and restore colons
+    if (modelId.startsWith(OLLAMA_MODEL_PREFIX)) {
+      const stripped = modelId.slice(OLLAMA_MODEL_PREFIX.length);
+      // Restore the last dash to colon for tag separator (e.g., llama3.1-8b → llama3.1:8b)
+      // Use cached model list for accurate mapping
+      if (this._modelCache) {
+        const cached = this._modelCache.find(m => m.name === modelId);
+        if (cached) return cached.ollamaName;
+      }
+      // Fallback: restore last dash before a version-like segment to colon
+      return stripped.replace(/-(\d+[bgBG]?)$/, ':$1')
+                     .replace(/-latest$/, ':latest');
+    }
+    return modelId;
+  }
+
+  /**
+   * Estimate token count from messages (fallback when Ollama doesn't report)
+   * @private
+   */
+  _estimateTokens(messages) {
+    let chars = 0;
+    for (const msg of messages) {
+      const content = msg.content || '';
+      chars += typeof content === 'string' ? content.length : JSON.stringify(content).length;
+    }
+    return Math.ceil(chars / 3.5); // ~3.5 chars per token average
+  }
+
+  /**
+   * Check if a model ID is an Ollama model
+   * @param {string} modelId - Model identifier
+   * @returns {boolean}
+   */
+  static isOllamaModel(modelId) {
+    return typeof modelId === 'string' && modelId.startsWith(OLLAMA_MODEL_PREFIX);
+  }
+
+  /**
+   * Pull a model from Ollama registry
+   * @param {string} modelName - Model to pull (e.g., "llama3.1:8b")
+   * @param {Function} onProgress - Progress callback
+   */
+  async pullModel(modelName, onProgress = null) {
+    this.logger?.info(`[Ollama] Pulling model: ${modelName}`);
+
+    try {
+      const stream = await this.client.pull({ model: modelName, stream: true });
+
+      if (!stream) {
+        throw new Error(`Ollama returned no stream when pulling "${modelName}". Check that Ollama is running.`);
+      }
+
+      for await (const progress of stream) {
+        if (onProgress) {
+          onProgress({
+            status: progress.status,
+            total: progress.total,
+            completed: progress.completed,
+            percent: progress.total ? Math.round((progress.completed / progress.total) * 100) : null
+          });
+        }
+      }
+
+      // Invalidate model cache
+      this._modelCache = null;
+      this.logger?.info(`[Ollama] Successfully pulled ${modelName}`);
+      return { success: true, model: modelName };
+    } catch (err) {
+      this.logger?.error(`[Ollama] Pull failed for ${modelName}: ${err.message}`);
+      throw err;
+    }
+  }
+
+  /**
+   * Delete a model from Ollama
+   * @param {string} modelName - Model to delete
+   */
+  async deleteModel(modelName) {
+    try {
+      await this.client.delete({ model: modelName });
+      this._modelCache = null;
+      this.logger?.info(`[Ollama] Deleted model: ${modelName}`);
+      return { success: true };
+    } catch (err) {
+      this.logger?.error(`[Ollama] Delete failed for ${modelName}: ${err.message}`);
+      throw err;
+    }
+  }
+
+  /**
+   * Get model info from Ollama
+   * @param {string} modelName - Model name
+   */
+  async getModelInfo(modelName) {
+    try {
+      const info = await this.client.show({ model: modelName });
+      return {
+        success: true,
+        info: {
+          name: modelName,
+          license: info.license,
+          modelfile: info.modelfile,
+          parameters: info.parameters,
+          template: info.template,
+          details: info.details
+        }
+      };
+    } catch (err) {
+      return { success: false, error: err.message };
+    }
+  }
+}
+
+// Singleton
+let instance = null;
+
+export function getOllamaService(config = {}, logger = null) {
+  if (!instance) {
+    instance = new OllamaService(config, logger);
+  }
+  return instance;
+}
+
+export { OllamaService, OLLAMA_MODEL_PREFIX };
+export default OllamaService;