KAMAL18/dual_agent_simulation
0
1import streamlit as st2import pandas as pd3import os4from huggingface_hub import InferenceClient5 6# Initialize HF Inference client7@st.cache_resource8def get_client():9 # Try to get token from environment variables (works in Hugging Face Spaces)10 hf_token = os.environ.get("HF_TOKEN")11 if not hf_token:12 try:13 hf_token = st.secrets["HF_TOKEN"] # Fallback to Streamlit secrets14 except:15 st.error("HF_TOKEN not found. Please set it in your Space's settings or secrets.toml")16 st.stop()17 return InferenceClient(token=hf_token)18 19client = get_client()20 21# Define the LLM call function22def call_llama(prompt, model="mistralai/Mistral-7B-Instruct-v0.2", max_tokens=500):23 try:24 response = client.text_generation(25 prompt=prompt,26 model=model,27 max_new_tokens=max_tokens,28 temperature=0.729 )30 return response31 except Exception as e:32 st.error(f"Error calling LLM: {e}")33 return "Sorry, I encountered an error."34 35# Set page config36st.set_page_config(layout="wide")37st.title("Duel Agent Simulation 🦙🦙")38 39# Sidebar setup40with st.sidebar:41 with st.expander("Instruction Manual"):42 st.markdown("""43 # 🦙🦙 Duel Agent Simulation44 ## Overview45 This app simulates an interview with two AI agents:46 1. **Interviewer**: Asks questions about your topic47 2. **Interviewee**: Attempts to answer (poorly at first)48 3. **Judge**: Provides feedback after each answer49 50 ## How to Use51 1. Enter a topic below52 2. Click "Run Simulation"53 3. Watch the conversation unfold54 4. See the evaluation results55 56 The simulation stops when the interviewee gives a good answer (8/10 or higher).57 """)58 59 # User inputs60 user_topic = st.text_input("Enter a topic", "Artificial Intelligence")61 submit_button = st.button("Run Simulation!")62 63 if st.button("Clear Session"):64 st.session_state.clear()65 st.rerun()66 67# Initialize session state68if "messages" not in st.session_state:69 st.session_state.messages = []70 71if "simulation_data" not in st.session_state:72 st.session_state.simulation_data = {73 "iterations": [],74 "questions": [],75 "answers": [],76 "feedback": [],77 "scores": []78 }79 80# Display chat history81for message in st.session_state.messages:82 with st.chat_message(message["role"]):83 st.markdown(message["content"])84 85# Run simulation when button is pressed86if submit_button:87 iter_count = 088 current_prompt = f"Ask a technical interview question about: {user_topic}"89 90 # Display initial topic91 with st.chat_message("user"):92 st.markdown(f"**Topic:** {user_topic}")93 st.session_state.messages.append({"role": "user", "content": f"Topic: {user_topic}"})94 95 with st.spinner("Running simulation..."):96 while iter_count < 6: # Max 6 iterations97 # Interviewer asks question98 question = call_llama(99 f"""You are a technical interviewer. Ask one specific question about {user_topic}.100 Make it challenging but answerable. Return only the question."""101 ).strip()102 103 with st.chat_message("assistant"):104 st.markdown(f"**Interviewer:** {question}")105 st.session_state.messages.append({"role": "assistant", "content": f"Interviewer: {question}"})106 107 # Interviewee answers108 if iter_count < 2: # First attempts are poor109 answer_prompt = f"""You are a nervous interviewee. Give a mediocre answer to:110 "{question}". Return only the answer."""111 else: # Later attempts improve112 feedback = st.session_state.simulation_data["feedback"][-1] if iter_count > 0 else ""113 answer_prompt = f"""You're learning to answer better. Previous feedback was:114 "{feedback}". Now answer: "{question}". Return only the improved answer."""115 116 answer = call_llama(answer_prompt).strip()117 118 with st.chat_message("user"):119 st.markdown(f"**Interviewee:** {answer}")120 st.session_state.messages.append({"role": "user", "content": f"Interviewee: {answer}"})121 122 # Judge evaluates123 feedback_prompt = f"""Evaluate this interview exchange:124 Question: {question}125 Answer: {answer}126 Provide specific feedback and a score from 1-10 (10=best). 127 Format: Feedback: [your feedback] Score: [1-10]"""128 129 judge_response = call_llama(feedback_prompt).strip()130 131 # Extract score132 score = 5 # default133 if "Score:" in judge_response:134 try:135 score_part = judge_response.split("Score:")[1].strip()136 score = int(score_part.split()[0])137 except (ValueError, IndexError):138 pass139 140 # Store data141 st.session_state.simulation_data["iterations"].append(iter_count)142 st.session_state.simulation_data["questions"].append(question)143 st.session_state.simulation_data["answers"].append(answer)144 st.session_state.simulation_data["feedback"].append(judge_response)145 st.session_state.simulation_data["scores"].append(score)146 147 # Show feedback148 with st.chat_message("assistant"):149 st.markdown(f"**Judge:** {judge_response}")150 st.session_state.messages.append({"role": "assistant", "content": f"Judge: {judge_response}"})151 152 # Display results table153 with st.expander("Detailed Results"):154 results_df = pd.DataFrame({155 "Round": st.session_state.simulation_data["iterations"],156 "Question": st.session_state.simulation_data["questions"],157 "Answer": st.session_state.simulation_data["answers"],158 "Score": st.session_state.simulation_data["scores"],159 "Feedback": st.session_state.simulation_data["feedback"]160 })161 st.dataframe(results_df, use_container_width=True)162 163 # Check stopping condition164 if score >= 8:165 st.success("🎉 Simulation complete! The interviewee passed with a good answer.")166 break167 168 iter_count += 1169 current_prompt = f"Ask a follow-up question about: {user_topic}"170 171 if iter_count == 6:172 st.warning("Simulation ended - maximum rounds reached")