ak-gemini 1.1.13 → 1.2.0

package/agent.js

@@ -0,0 +1,481 @@
+/**
+ * @fileoverview AIAgent class for chat-based agent interactions with built-in tools.
+ * Supports streaming and non-streaming conversations with HTTP and markdown tools.
+ */
+
+import dotenv from 'dotenv';
+dotenv.config();
+const { NODE_ENV = "unknown", LOG_LEVEL = "" } = process.env;
+
+import { GoogleGenAI, HarmCategory, HarmBlockThreshold, ThinkingLevel } from '@google/genai';
+import log from './logger.js';
+import { BUILT_IN_DECLARATIONS, executeBuiltInTool } from './tools.js';
+
+const DEFAULT_SAFETY_SETTINGS = [
+	{ category: HarmCategory.HARM_CATEGORY_HARASSMENT, threshold: HarmBlockThreshold.BLOCK_NONE },
+	{ category: HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT, threshold: HarmBlockThreshold.BLOCK_NONE }
+];
+
+const DEFAULT_THINKING_CONFIG = {
+	thinkingBudget: 0
+};
+
+const THINKING_SUPPORTED_MODELS = [
+	/^gemini-3-flash(-preview)?$/,
+	/^gemini-3-pro(-preview|-image-preview)?$/,
+	/^gemini-2\.5-pro/,
+	/^gemini-2\.5-flash(-preview)?$/,
+	/^gemini-2\.5-flash-lite(-preview)?$/,
+	/^gemini-2\.0-flash$/
+];
+
+/**
+ * @typedef {import('./types').AIAgentOptions} AIAgentOptions
+ * @typedef {import('./types').AgentResponse} AgentResponse
+ * @typedef {import('./types').AgentStreamEvent} AgentStreamEvent
+ * @typedef {import('./types').UsageData} UsageData
+ */
+
+/**
+ * Chat-based AI agent with built-in tools for HTTP requests and markdown generation.
+ *
+ * Unlike AITransformer (which is optimized for few-shot JSON transformations), AIAgent
+ * is designed for interactive, multi-turn conversations where the agent can take actions
+ * on the user's behalf — like fetching data from APIs, posting to endpoints, and
+ * generating structured markdown reports.
+ *
+ * Built-in tools:
+ * - **http_get** — Fetch any URL (APIs, web pages, etc.)
+ * - **http_post** — POST JSON to any endpoint
+ * - **write_markdown** — Generate markdown documents (reports, summaries, findings)
+ *
+ * The agent automatically manages the tool-use loop: when the model decides to call
+ * a tool, the agent executes it, sends the result back, and continues until the model
+ * produces a final text response.
+ *
+ * @example
+ * ```javascript
+ * import { AIAgent } from 'ak-gemini';
+ *
+ * const agent = new AIAgent({
+ *   systemPrompt: 'You are a research assistant...',
+ *   onMarkdown: (filename, content) => fs.writeFileSync(filename, content)
+ * });
+ *
+ * // Non-streaming
+ * const res = await agent.chat('Fetch https://api.example.com/data and summarize it');
+ * console.log(res.text);          // Agent's response
+ * console.log(res.toolCalls);     // [{name, args, result}, ...]
+ * console.log(res.markdownFiles); // [{filename, content}, ...]
+ *
+ * // Streaming
+ * for await (const event of agent.stream('Write a report on...')) {
+ *   if (event.type === 'text') process.stdout.write(event.text);
+ *   if (event.type === 'tool_call') console.log(`Calling: ${event.toolName}`);
+ *   if (event.type === 'done') console.log('\nDone!');
+ * }
+ * ```
+ */
+class AIAgent {
+	/**
+	 * Create a new AIAgent instance.
+	 * @param {AIAgentOptions} [options={}] - Configuration options (see AIAgentOptions in types.d.ts)
+	 */
+	constructor(options = {}) {
+		this.modelName = options.modelName || 'gemini-2.5-flash';
+		this.systemPrompt = options.systemPrompt || 'You are a helpful AI assistant.';
+		this.maxToolRounds = options.maxToolRounds || 10;
+		this.httpTimeout = options.httpTimeout || 30000;
+		this.maxRetries = options.maxRetries || 3;
+		this.onToolCall = options.onToolCall || null;
+		this.onMarkdown = options.onMarkdown || null;
+		this.labels = options.labels || {};
+
+		// Auth - same as AITransformer
+		this.vertexai = options.vertexai || false;
+		this.project = options.project || process.env.GOOGLE_CLOUD_PROJECT || null;
+		this.location = options.location || process.env.GOOGLE_CLOUD_LOCATION || undefined;
+		this.googleAuthOptions = options.googleAuthOptions || null;
+		this.apiKey = options.apiKey !== undefined && options.apiKey !== null ? options.apiKey : process.env.GEMINI_API_KEY;
+
+		if (!this.vertexai && !this.apiKey) {
+			throw new Error("Missing Gemini API key. Provide via options.apiKey or GEMINI_API_KEY env var. For Vertex AI, set vertexai: true with project and location.");
+		}
+		if (this.vertexai && !this.project) {
+			throw new Error("Vertex AI requires a project ID. Provide via options.project or GOOGLE_CLOUD_PROJECT env var.");
+		}
+
+		// Log level
+		this._configureLogLevel(options.logLevel);
+
+		// Build chat config
+		this.chatConfig = {
+			temperature: 0.7,
+			topP: 0.95,
+			topK: 64,
+			safetySettings: DEFAULT_SAFETY_SETTINGS,
+			systemInstruction: this.systemPrompt,
+			maxOutputTokens: options.chatConfig?.maxOutputTokens || 50_000,
+			...options.chatConfig
+		};
+
+		// Ensure systemPrompt takes precedence over chatConfig.systemInstruction
+		this.chatConfig.systemInstruction = this.systemPrompt;
+
+		// Thinking config
+		this._configureThinking(options.thinkingConfig);
+
+		// Tools config
+		this.chatConfig.tools = [{ functionDeclarations: BUILT_IN_DECLARATIONS }];
+		this.chatConfig.toolConfig = { functionCallingConfig: { mode: 'AUTO' } };
+
+		// State
+		this.genAIClient = null;
+		this.chatSession = null;
+		this.lastResponseMetadata = null;
+		this._markdownFiles = [];
+
+		log.debug(`AIAgent created with model: ${this.modelName}`);
+	}
+
+	/**
+	 * Initialize the agent — creates the GenAI client and chat session.
+	 * Called automatically by chat() and stream() if not called explicitly.
+	 * Idempotent — safe to call multiple times.
+	 * @returns {Promise<void>}
+	 */
+	async init() {
+		if (this.chatSession) return;
+
+		const clientOptions = this.vertexai
+			? {
+				vertexai: true,
+				project: this.project,
+				...(this.location && { location: this.location }),
+				...(this.googleAuthOptions && { googleAuthOptions: this.googleAuthOptions })
+			}
+			: { apiKey: this.apiKey };
+
+		this.genAIClient = new GoogleGenAI(clientOptions);
+
+		this.chatSession = this.genAIClient.chats.create({
+			model: this.modelName,
+			config: {
+				...this.chatConfig,
+				...(this.vertexai && Object.keys(this.labels).length > 0 && { labels: this.labels })
+			},
+			history: []
+		});
+
+		try {
+			await this.genAIClient.models.list();
+			log.debug("AIAgent: Gemini API connection successful.");
+		} catch (e) {
+			throw new Error(`AIAgent initialization failed: ${e.message}`);
+		}
+
+		log.debug("AIAgent: Chat session initialized.");
+	}
+
+	/**
+	 * Send a message and get a complete response (non-streaming).
+	 * Automatically handles the tool-use loop — if the model requests tool calls,
+	 * they are executed and results sent back until the model produces a final response.
+	 *
+	 * @param {string} message - The user's message
+	 * @returns {Promise<AgentResponse>} Response with text, toolCalls, markdownFiles, and usage
+	 * @example
+	 * const res = await agent.chat('Fetch https://api.example.com/users');
+	 * console.log(res.text);      // Agent's summary
+	 * console.log(res.toolCalls); // [{name: 'http_get', args: {...}, result: {...}}]
+	 */
+	async chat(message) {
+		if (!this.chatSession) await this.init();
+
+		this._markdownFiles = [];
+		const allToolCalls = [];
+
+		let response = await this.chatSession.sendMessage({ message });
+
+		for (let round = 0; round < this.maxToolRounds; round++) {
+			const functionCalls = response.functionCalls;
+			if (!functionCalls || functionCalls.length === 0) break;
+
+			// Execute all tool calls in parallel
+			const toolResults = await Promise.all(
+				functionCalls.map(async (call) => {
+					let result;
+					try {
+						result = await executeBuiltInTool(call.name, call.args, {
+							httpTimeout: this.httpTimeout,
+							onToolCall: this.onToolCall,
+							onMarkdown: this.onMarkdown
+						});
+					} catch (err) {
+						log.warn(`Tool ${call.name} failed: ${err.message}`);
+						result = { error: err.message };
+					}
+
+					allToolCalls.push({ name: call.name, args: call.args, result });
+
+					// Collect markdown files
+					if (call.name === 'write_markdown' && call.args) {
+						this._markdownFiles.push({
+							filename: /** @type {string} */ (call.args.filename),
+							content: /** @type {string} */ (call.args.content)
+						});
+					}
+
+					return { id: call.id, name: call.name, result };
+				})
+			);
+
+			// Send all function responses back to the model
+			response = await this.chatSession.sendMessage({
+				message: toolResults.map(r => ({
+					functionResponse: {
+						id: r.id,
+						name: r.name,
+						response: { output: r.result }
+					}
+				}))
+			});
+		}
+
+		// Capture metadata
+		this._captureMetadata(response);
+
+		return {
+			text: response.text || '',
+			toolCalls: allToolCalls,
+			markdownFiles: [...this._markdownFiles],
+			usage: this.getLastUsage()
+		};
+	}
+
+	/**
+	 * Send a message and stream the response as events.
+	 * Automatically handles the tool-use loop between streamed rounds.
+	 *
+	 * Event types:
+	 * - `text` — A chunk of the agent's text response (yield as it arrives)
+	 * - `tool_call` — The agent is about to call a tool (includes toolName and args)
+	 * - `tool_result` — A tool finished executing (includes toolName and result)
+	 * - `markdown` — A markdown document was generated (includes filename and content)
+	 * - `done` — The agent finished (includes fullText, markdownFiles, usage)
+	 *
+	 * @param {string} message - The user's message
+	 * @yields {AgentStreamEvent}
+	 * @example
+	 * for await (const event of agent.stream('Analyze this API...')) {
+	 *   if (event.type === 'text') process.stdout.write(event.text);
+	 *   if (event.type === 'tool_call') console.log(`Calling: ${event.toolName}`);
+	 *   if (event.type === 'done') console.log(`\nTokens: ${event.usage?.totalTokens}`);
+	 * }
+	 */
+	async *stream(message) {
+		if (!this.chatSession) await this.init();
+
+		this._markdownFiles = [];
+		const allToolCalls = [];
+		let fullText = '';
+
+		let streamResponse = await this.chatSession.sendMessageStream({ message });
+
+		for (let round = 0; round < this.maxToolRounds; round++) {
+			let roundText = '';
+			const functionCalls = [];
+
+			// Consume the stream
+			for await (const chunk of streamResponse) {
+				// Check for function calls first (accessing .text when functionCall parts
+				// exist triggers a warning from the SDK)
+				if (chunk.functionCalls) {
+					functionCalls.push(...chunk.functionCalls);
+				} else if (chunk.candidates?.[0]?.content?.parts?.[0]?.text) {
+					const text = chunk.candidates[0].content.parts[0].text;
+					roundText += text;
+					fullText += text;
+					yield { type: 'text', text };
+				}
+			}
+
+			// No tool calls — we're done
+			if (functionCalls.length === 0) {
+				// Capture metadata from the last chunk's aggregated response
+				// (streaming doesn't give us a final response object easily,
+				// so metadata may be limited here)
+				yield {
+					type: 'done',
+					fullText,
+					markdownFiles: [...this._markdownFiles],
+					usage: this.getLastUsage()
+				};
+				return;
+			}
+
+			// Execute tools sequentially so we can yield events
+			const toolResults = [];
+			for (const call of functionCalls) {
+				yield { type: 'tool_call', toolName: call.name, args: call.args };
+
+				let result;
+				try {
+					result = await executeBuiltInTool(call.name, call.args, {
+						httpTimeout: this.httpTimeout,
+						onToolCall: this.onToolCall,
+						onMarkdown: this.onMarkdown
+					});
+				} catch (err) {
+					log.warn(`Tool ${call.name} failed: ${err.message}`);
+					result = { error: err.message };
+				}
+
+				allToolCalls.push({ name: call.name, args: call.args, result });
+				yield { type: 'tool_result', toolName: call.name, result };
+
+				if (call.name === 'write_markdown' && call.args) {
+					const mdFilename = /** @type {string} */ (call.args.filename);
+					const mdContent = /** @type {string} */ (call.args.content);
+					this._markdownFiles.push({ filename: mdFilename, content: mdContent });
+					yield { type: 'markdown', filename: mdFilename, content: mdContent };
+				}
+
+				toolResults.push({ id: call.id, name: call.name, result });
+			}
+
+			// Send function responses back and get next stream
+			streamResponse = await this.chatSession.sendMessageStream({
+				message: toolResults.map(r => ({
+					functionResponse: {
+						id: r.id,
+						name: r.name,
+						response: { output: r.result }
+					}
+				}))
+			});
+		}
+
+		// Max rounds reached
+		yield {
+			type: 'done',
+			fullText,
+			markdownFiles: [...this._markdownFiles],
+			usage: this.getLastUsage(),
+			warning: 'Max tool rounds reached'
+		};
+	}
+
+	/**
+	 * Clear conversation history while preserving tools and system prompt.
+	 * Useful for starting a new user session without re-initializing the agent.
+	 * @returns {Promise<void>}
+	 */
+	async clearHistory() {
+		this.chatSession = this.genAIClient.chats.create({
+			model: this.modelName,
+			config: {
+				...this.chatConfig,
+				...(this.vertexai && Object.keys(this.labels).length > 0 && { labels: this.labels })
+			},
+			history: []
+		});
+		this._markdownFiles = [];
+		this.lastResponseMetadata = null;
+		log.debug("AIAgent: Conversation history cleared.");
+	}
+
+	/**
+	 * Get conversation history.
+	 * @param {boolean} [curated=false]
+	 * @returns {any[]}
+	 */
+	getHistory(curated = false) {
+		if (!this.chatSession) return [];
+		return this.chatSession.getHistory(curated);
+	}
+
+	/**
+	 * Get structured usage data from the last API call.
+	 * Returns null if no API call has been made yet.
+	 * @returns {UsageData|null} Usage data with promptTokens, responseTokens, totalTokens, etc.
+	 */
+	getLastUsage() {
+		if (!this.lastResponseMetadata) return null;
+		const m = this.lastResponseMetadata;
+		return {
+			promptTokens: m.promptTokens,
+			responseTokens: m.responseTokens,
+			totalTokens: m.totalTokens,
+			attempts: 1,
+			modelVersion: m.modelVersion,
+			requestedModel: this.modelName,
+			timestamp: m.timestamp
+		};
+	}
+
+	// --- Private helpers ---
+
+	/**
+	 * Capture response metadata (model version, token counts) from an API response.
+	 * @param {import('@google/genai').GenerateContentResponse} response
+	 * @private
+	 */
+	_captureMetadata(response) {
+		this.lastResponseMetadata = {
+			modelVersion: response.modelVersion || null,
+			requestedModel: this.modelName,
+			promptTokens: response.usageMetadata?.promptTokenCount || 0,
+			responseTokens: response.usageMetadata?.candidatesTokenCount || 0,
+			totalTokens: response.usageMetadata?.totalTokenCount || 0,
+			timestamp: Date.now()
+		};
+	}
+
+	/** @private */
+	_configureLogLevel(logLevel) {
+		if (logLevel) {
+			if (logLevel === 'none') {
+				log.level = 'silent';
+			} else {
+				log.level = logLevel;
+			}
+		} else if (LOG_LEVEL) {
+			log.level = LOG_LEVEL;
+		} else if (NODE_ENV === 'dev') {
+			log.level = 'debug';
+		} else if (NODE_ENV === 'test') {
+			log.level = 'warn';
+		} else if (NODE_ENV.startsWith('prod')) {
+			log.level = 'error';
+		} else {
+			log.level = 'info';
+		}
+	}
+
+	/** @private */
+	_configureThinking(thinkingConfig) {
+		const modelSupportsThinking = THINKING_SUPPORTED_MODELS.some(p => p.test(this.modelName));
+
+		if (thinkingConfig === undefined) return;
+
+		if (thinkingConfig === null) {
+			delete this.chatConfig.thinkingConfig;
+			return;
+		}
+
+		if (!modelSupportsThinking) {
+			log.warn(`Model ${this.modelName} does not support thinking features. Ignoring thinkingConfig.`);
+			return;
+		}
+
+		const config = { ...DEFAULT_THINKING_CONFIG, ...thinkingConfig };
+		if (thinkingConfig.thinkingLevel !== undefined) {
+			delete config.thinkingBudget;
+		}
+		this.chatConfig.thinkingConfig = config;
+		log.debug(`Thinking config applied: ${JSON.stringify(config)}`);
+	}
+}
+
+export default AIAgent;