from openai import OpenAI import json import os import requests from PyPDF2 import PdfReader import gradio as gr from pydantic import BaseModel def push(text): requests.post( "https://api.pushover.net/1/messages.json", data={ "token": os.getenv("PUSHOVER_TOKEN"), "user": os.getenv("PUSHOVER_USER"), "message": text, } ) def record_user_details(email, name="Name not provided", notes="not provided"): push(f"Recording {name} with email {email} and notes {notes}") return {"recorded": "ok"} def record_unknown_question(question): push(f"Recording {question}") return {"recorded": "ok"} record_user_details_json = { "name": "record_user_details", "description": "Use this tool to record that a user is interested in being in touch and provided an email address", "parameters": { "type": "object", "properties": { "email": { "type": "string", "description": "The email address of this user" }, "name": { "type": "string", "description": "The user's name, if they provided it" } , "notes": { "type": "string", "description": "Any additional information about the conversation that's worth recording to give context" } }, "required": ["email"], "additionalProperties": False } } record_unknown_question_json = { "name": "record_unknown_question", "description": "Always use this tool to record any question that couldn't be answered as you didn't know the answer", "parameters": { "type": "object", "properties": { "question": { "type": "string", "description": "The question that couldn't be answered" }, }, "required": ["question"], "additionalProperties": False } } tools = [{"type": "function", "function": record_user_details_json}, {"type": "function", "function": record_unknown_question_json}] # Create a Pydantic model for the Evaluation class Evaluation(BaseModel): is_acceptable: bool feedback: str class Me: def __init__(self): # when saving secret in HF space, don't use "" :-) # Initialize Open Router client using OpenAI format # open_router_api_key = os.getenv('OPEN_ROUTER_API_KEY') # if open_router_api_key: # print(f"Checking Keys: Open router API Key exists and begins {open_router_api_key[:8]}") # else: # print("Checking Keys: Open router API Key not set - please head to the troubleshooting guide in the setup folder") self.openrouter = OpenAI( base_url="https://openrouter.ai/api/v1", api_key= os.getenv('OPEN_ROUTER_API_KEY') ) # open_router_api_key # Initialize Gemini client using OpenAI format self.gemini = OpenAI( api_key=os.getenv("GOOGLE_API_KEY"), base_url="https://generativelanguage.googleapis.com/v1beta/openai/" ) self.name = "Chaoran Zhou" reader = PdfReader("me/linkedin.pdf") self.linkedin = "" for page in reader.pages: text = page.extract_text() if text: self.linkedin += text with open("me/summary.txt", "r", encoding="utf-8") as f: self.summary = f.read() def handle_tool_call(self, tool_calls): results = [] for tool_call in tool_calls: tool_name = tool_call.function.name arguments = json.loads(tool_call.function.arguments) print(f"Tool called: {tool_name}", flush=True) tool = globals().get(tool_name) result = tool(**arguments) if tool else {} results.append({"role": "tool","content": json.dumps(result),"tool_call_id": tool_call.id}) return results def system_prompt(self): system_prompt = f"You are acting as {self.name}. You are answering questions on {self.name}'s website, \ particularly questions related to {self.name}'s career, background, skills and experience. \ Your responsibility is to represent {self.name} for interactions on the website as faithfully as possible. \ You are given a summary of {self.name}'s background and LinkedIn profile which you can use to answer questions. \ Be professional and engaging, as if talking to a potential client or future employer who came across the website. \ If you don't know the answer to any question, use your record_unknown_question tool to record the question that you couldn't answer, even if it's about something trivial or unrelated to career. \ If the user is engaging in discussion, try to steer them towards getting in touch via email; ask for their email and record it using your record_user_details tool. " system_prompt += f"\n\n## Summary:\n{self.summary}\n\n## LinkedIn Profile:\n{self.linkedin}\n\n" system_prompt += f"With this context, please chat with the user, always staying in character as {self.name}." return system_prompt def evaluator_system_prompt(self): evaluator_system_prompt = f"You are an evaluator that decides whether a response to a question is acceptable. \ You are provided with a conversation between a User and an Agent. Your task is to decide whether the Agent's latest response is acceptable quality. \ The Agent is playing the role of {self.name} and is representing {self.name} on their website. \ The Agent has been instructed to be professional and engaging, as if talking to a potential client or future employer who came across the website. \ The Agent has been provided with context on {self.name} in the form of their summary and LinkedIn details. Here's the information:" evaluator_system_prompt += f"\n\n## Summary:\n{self.summary}\n\n## LinkedIn Profile:\n{self.linkedin}\n\n" evaluator_system_prompt += f"With this context, please evaluate the latest response, replying with whether the response is acceptable and your feedback." return evaluator_system_prompt def evaluator_user_prompt(self, reply, message, history): user_prompt = f"Here's the conversation between the User and the Agent: \n\n{history}\n\n" user_prompt += f"Here's the latest message from the User: \n\n{message}\n\n" user_prompt += f"Here's the latest response from the Agent: \n\n{reply}\n\n" user_prompt += f"Please evaluate the response, replying with whether it is acceptable and your feedback." return user_prompt def evaluate(self, reply, message, history) -> Evaluation: messages = [ {"role": "system", "content": self.evaluator_system_prompt()}, {"role": "user", "content": self.evaluator_user_prompt(reply, message, history)} ] response = self.gemini.beta.chat.completions.parse( model="gemini-2.5-flash-preview-05-20", messages=messages, response_format=Evaluation ) return response.choices[0].message.parsed def rerun(self, reply, message, history, feedback): updated_system_prompt = self.system_prompt() + f"\n\n## Previous answer rejected\nYou just tried to reply, but the quality control rejected your reply\n" updated_system_prompt += f"## Your attempted answer:\n{reply}\n\n" updated_system_prompt += f"## Reason for rejection:\n{feedback}\n\n" messages = [{"role": "system", "content": updated_system_prompt}] + history + [{"role": "user", "content": message}] done = False while not done: response = self.gemini.chat.completions.create( model="gemini-2.5-flash-preview-05-20", messages=messages, tools=tools ) if response.choices[0].finish_reason == "tool_calls": message_obj = response.choices[0].message tool_calls = message_obj.tool_calls results = self.handle_tool_call(tool_calls) messages.append(message_obj) messages.extend(results) else: done = True return response.choices[0].message.content def chat(self, message, history): messages = [{"role": "system", "content": self.system_prompt()}] + history + [{"role": "user", "content": message}] done = False # Generate initial response with tool handling while not done: response = self.openrouter.chat.completions.create(model="meta-llama/llama-3.3-8b-instruct:free", messages=messages, tools=tools) if response.choices[0].finish_reason=="tool_calls": message = response.choices[0].message tool_calls = message.tool_calls results = self.handle_tool_call(tool_calls) messages.append(message) messages.extend(results) else: done = True reply = response.choices[0].message.content # Evaluate the response try: evaluation = self.evaluate(reply, message, history) if evaluation.is_acceptable: print("Passed evaluation - returning reply") else: print("Failed evaluation - retrying") print(f"Feedback: {evaluation.feedback}") reply = self.rerun(reply, message, history, evaluation.feedback) except Exception as e: print(f"Evaluation failed with error: {e}") print("Proceeding with original reply") return reply if __name__ == "__main__": me = Me() gr.ChatInterface(me.chat, type="messages").launch(debug=True, share=False)