feat: solution for '8. Самописный поисковый агент на основе deep agents from scratch'

src/deepAgent.js

@@ -0,0 +1,159 @@
+const { OpenAI } = require('langchain/llms/openai');
+const { Tool } = require('langchain/tools');
+const { PromptTemplate } = require('langchain/prompts');
+const { ConversationBufferMemory } = require('langchain/memory');
+
+/**
+ * Planner class that uses an LLM to generate a plan of tool calls.
+ */
+class Planner {
+  /**
+   * @param {OpenAI} llm - The language model to use for planning.
+   * @param {Tool[]} tools - Available tools for the agent.
+   */
+  constructor(llm, tools) {
+    this.llm = llm;
+    this.tools = tools;
+
+    // Prompt template for the planner.
+    this.plannerPrompt = new PromptTemplate({
+      template: `You are a planning assistant. Given the user query: "{input}", produce a JSON array of steps. Each step must contain:
+- "tool": the name of the tool to use (must be one of: ${tools.map(t => t.name).join(', ')}),
+- "input": the input string for that tool.
+
+The JSON array should be the only output. Example:
+[
+  {"tool":"Echo","input":"Hello"}
+]`,
+      inputVariables: ['input'],
+    });
+  }
+
+  /**
+   * Generates a plan for the given input.
+   * @param {string} input - The user query.
+   * @returns {Promise<Array<{tool: string, input: string}>>} The plan steps.
+   */
+  async plan(input) {
+    const prompt = this.plannerPrompt.format({ input });
+    const raw = await this.llm.invoke(prompt);
+    let plan;
+    try {
+      plan = JSON.parse(raw);
+      if (!Array.isArray(plan)) throw new Error('Plan is not an array');
+    } catch (e) {
+      throw new Error(`Planner failed to parse JSON: ${e.message}. Raw output: ${raw}`);
+    }
+    // Validate tool names
+    for (const step of plan) {
+      if (!this.tools.find(t => t.name === step.tool)) {
+        throw new Error(`Planner suggested unknown tool "${step.tool}"`);
+      }
+    }
+    return plan;
+  }
+}
+
+/**
+ * Executor class that runs the planned tool calls sequentially.
+ */
+class Executor {
+  /**
+   * @param {Tool[]} tools - Available tools for the agent.
+   */
+  constructor(tools) {
+    this.tools = tools;
+  }
+
+  /**
+   * Executes the plan and returns the final output.
+   * @param {Array<{tool: string, input: string}>} plan - The plan steps.
+   * @returns {Promise<string>} The final output after executing all steps.
+   */
+  async execute(plan) {
+    let lastOutput = '';
+    for (const step of plan) {
+      const tool = this.tools.find(t => t.name === step.tool);
+      if (!tool) {
+        throw new Error(`Executor cannot find tool "${step.tool}"`);
+      }
+      const output = await tool.func(step.input);
+      lastOutput = output;
+    }
+    return lastOutput;
+  }
+}
+
+/**
+ * DeepAgent implements the Deep Agents from Scratch template.
+ */
+class DeepAgent {
+  /**
+   * @param {Object} options Configuration options.
+   * @param {string} [options.modelName='gpt-3.5-turbo'] The LLM model to use.
+   * @param {number} [options.temperature=0.7] Temperature for LLM sampling.
+   * @param {number} [options.maxTokens=512] Maximum tokens for LLM output.
+   * @param {string} [options.apiKey] OpenAI API key. If not provided, will use process.env.OPENAI_API_KEY.
+   */
+  constructor({
+    modelName = 'gpt-3.5-turbo',
+    temperature = 0.7,
+    maxTokens = 512,
+    apiKey,
+  } = {}) {
+    this.llm = new OpenAI({
+      modelName,
+      temperature,
+      maxTokens,
+      openAIApiKey: apiKey || process.env.OPENAI_API_KEY,
+    });
+
+    // Define a simple Echo tool that returns the input back.
+    const echoTool = new Tool({
+      name: 'Echo',
+      description: 'Echoes the input back to the user.',
+      func: async (input) => input,
+    });
+
+    // Memory component required by the Deep Agents from Scratch template.
+    this.memory = new ConversationBufferMemory({
+      memoryKey: 'chat_history',
+      inputKey: 'input',
+      outputKey: 'output',
+    });
+
+    this.tools = [echoTool];
+
+    // Instantiate Planner and Executor.
+    this.planner = new Planner(this.llm, this.tools);
+    this.executor = new Executor(this.tools);
+  }
+
+  /**
+   * Runs the agent on a given question.
+   * @param {string} question The question to answer.
+   * @returns {Promise<string>} The agent's answer.
+   */
+  async run(question) {
+    try {
+      // Store the question in memory.
+      await this.memory.saveContext({ input: question }, { output: '' });
+
+      // Generate a plan.
+      const plan = await this.planner.plan(question);
+
+      // Execute the plan.
+      const result = await this.executor.execute(plan);
+
+      // Save the result in memory.
+      await this.memory.saveContext({ input: question }, { output: result });
+
+      return result.trim();
+    } catch (err) {
+      console.error('DeepAgent encountered an error:', err);
+      throw err;
+    }
+  }
+}
+
+module.exports = { DeepAgent };

src/index.js

@@ -0,0 +1,18 @@
+require('dotenv').config();
+const { DeepAgent } = require('./deepAgent');
+
+(async () => {
+  const agent = new DeepAgent({
+    modelName: process.env.OPENAI_MODEL || 'gpt-3.5-turbo',
+    temperature: 0.7,
+  });
+
+  const query = process.argv[2] || 'What is the capital of France?';
+  console.log(`Query: ${query}`);
+  try {
+    const answer = await agent.run(query);
+    console.log(`Answer: ${answer}`);
+  } catch (err) {
+    console.error('Error running DeepAgent:', err);
+  }
+})();