Upload OpenRouter_Agent.py

OpenRouter_Agent.py

@@ -0,0 +1,133 @@
+"""High level multi-agent system powered by OpenRouter models.
+
+This module sets up a manager agent that delegates tasks to specialized
+web and information agents.  It relies on the ``smolagent`` framework and
+OpenRouter API models for language generation and verification.
+"""
+
+from smolagents import (
+    CodeAgent,
+    VisitWebpageTool,
+    WebSearchTool,
+    WikipediaSearchTool,
+    PythonInterpreterTool,
+    FinalAnswerTool,
+    OpenAIServerModel,
+)
+from smolagents.utils import encode_image_base64, make_image_url
+from vision_tool import image_reasoning_tool
+import os
+
+OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
+if not OPENROUTER_API_KEY:
+    raise EnvironmentError("OPENROUTER_API_KEY environment variable not set")
+
+common = dict(
+    api_base="https://openrouter.ai/api/v1",
+    api_key=OPENROUTER_API_KEY,
+)
+
+
+class MultiAgentSystem:
+    """Coordinates specialized agents and their underlying models.
+
+    The system instantiates a ``web_agent`` for browsing and data collection,
+    an ``info_agent`` for computation and image reasoning, and a
+    ``manager_agent`` that plans tasks and verifies answers using several
+    OpenRouter models.
+    """
+    def __init__(self):
+        self.deepseek_model = OpenAIServerModel(
+            model_id="deepseek/deepseek-r1-0528:free",
+            max_tokens=8096,
+            **common,
+        )
+        self.qwen_model = OpenAIServerModel(
+            model_id="qwen/qwen-2.5-coder-32b-instruct:free",
+            max_tokens=8096,
+            **common,
+        )
+        self.gemini_model = OpenAIServerModel(
+            model_id="google/gemini-2.0-flash-exp:free",
+            max_tokens=8096,
+            **common,
+        )
+
+        self.web_agent = CodeAgent(
+            model_id=self.qwen_model,
+            tools=[WebSearchTool(), VisitWebpageTool(), WikipediaSearchTool()],
+            name="web_agent",
+            description=(
+                "You are a web browsing agent. Whenever the given {task} involves browsing "
+                "the web or a specific website such as Wikipedia or YouTube, you will use "
+                "the provided tools. For web-based factual and retrieval tasks, be as precise and source-reliable as possible."
+            ),
+            additional_authorized_imports=[
+                "markdownify",
+                "json",
+                "requests",
+                "urllib.request",
+                "urllib.parse",
+                "wikipedia-api",
+            ],
+            verbosity_level=0,
+            max_steps=10,
+        )
+
+        self.info_agent = CodeAgent(
+            model_id=self.qwen_model,
+            tools=[PythonInterpreterTool(), image_reasoning_tool],
+            name="info_agent",
+            description=(
+                "You are an agent tasked with cleaning, parsing, calculating information, and performing OCR if images are provided in the {task}. "
+                "You can also analyze images using a vision model. You handle all math, code, and data manipulation. Use numpy, math, and available libraries. "
+                "For image or chess tasks, use pytesseract, PIL, chess, or the image_reasoning_tool as required."
+            ),
+            additional_authorized_imports=[
+                "numpy",
+                "math",
+                "pytesseract",
+                "PIL",
+                "chess",
+            ],
+            max_tokens=8096,
+        )
+
+        self.manager_agent = CodeAgent(
+            model_id=self.deepseek_model,
+            tools=[FinalAnswerTool()],
+            managed_agents=[self.web_agent, self.info_agent],
+            name="manager_agent",
+            description=(
+                "You are the manager. Given a {task}, plan which agent to use: "
+                "If web data is needed, delegate to web_agent. If math, parsing, image reasoning, or code is needed, use info_agent. "
+                "After collecting outputs, optionally cross-validate and check correctness, then finalize and submit the best answer using FinalAnswerTool. "
+                "For each task, explicitly explain your planning steps and reasons for choosing which agent, and always prefer the most accurate and complete answer possible."
+            ),
+            additional_authorized_imports=[
+                "json",
+                "pandas",
+                "numpy",
+            ],
+            planning_interval=3,
+            verbosity_level=2,
+            max_tokens=8096,
+            final_answer_check=[self.check_reasoning],
+            max_steps=8,
+        )
+
+    def check_reasoning(self, final_answer, agent_memory):
+        model_id = self.gemini_model
+        verification_prompt = (
+            f"Here is a user-given task and the agent steps: {agent_memory.get_succinct_steps()}. "
+            f"The proposed final answer is: {final_answer}. "
+            "Please check that the reasoning process is correct: do they correctly answer the given task? "
+            "First list reasons why yes/no, then write your final decision: PASS in caps lock if it is satisfactory, FAIL if it is not."
+        )
+        output = model(verification_prompt)
+        print("Feedback: ", output)
+        if "FAIL" in output:
+            raise Exception(output)
+        return True
+
+