Update src/app_job_copy_1.py

src/app_job_copy_1.py

@@ -1,3 +1,585 @@
+# import streamlit as st
+# import pandas as pd
+# import json
+# import os
+# from pydantic import BaseModel, Field
+# from typing import List, Set, Dict, Any, Optional
+# import time
+# from langchain_openai import ChatOpenAI
+# from langchain_core.messages import HumanMessage
+# from langchain_core.prompts import ChatPromptTemplate
+# from langchain_core.output_parsers import StrOutputParser
+# from langchain_core.prompts import PromptTemplate
+# import gspread
+# from google.oauth2 import service_account
+# import tiktoken
+
+# st.set_page_config(
+#     page_title="Candidate Matching App",
+#     page_icon="👨‍💻🎯",
+#     layout="wide"
+# )
+
+# # Define pydantic model for structured output
+# class Shortlist(BaseModel):
+#     fit_score: float = Field(description="A score between 0 and 10 indicating how closely the candidate profile matches the job requirements.")
+#     candidate_name: str = Field(description="The name of the candidate.")
+#     candidate_url: str = Field(description="The URL of the candidate's LinkedIn profile.")
+#     candidate_summary: str = Field(description="A brief summary of the candidate's skills and experience along with its educational background.")
+#     candidate_location: str = Field(description="The location of the candidate.")
+#     justification: str = Field(description="Justification for the shortlisted candidate with the fit score")
+
+# # Function to calculate tokens
+# def calculate_tokens(text, model="gpt-4o-mini"):
+#     """Calculate the number of tokens in a given text for a specific model"""
+#     try:
+#         # Get the encoding for the model
+#         if "gpt-4" in model:
+#             encoding = tiktoken.encoding_for_model("gpt-4o-mini")
+#         elif "gpt-3.5" in model:
+#             encoding = tiktoken.encoding_for_model("gpt-3.5-turbo")
+#         else:
+#             encoding = tiktoken.get_encoding("cl100k_base")  # Default for newer models
+            
+#         # Encode the text and return the token count
+#         return len(encoding.encode(text))
+#     except Exception as e:
+#         # If there's an error, make a rough estimate (1 token ≈ 4 chars)
+#         return len(text) // 4
+
+# # Function to display token usage
+# def display_token_usage():
+#     """Display token usage statistics"""
+#     if 'total_input_tokens' not in st.session_state:
+#         st.session_state.total_input_tokens = 0
+#     if 'total_output_tokens' not in st.session_state:
+#         st.session_state.total_output_tokens = 0
+    
+#     total_input = st.session_state.total_input_tokens
+#     total_output = st.session_state.total_output_tokens
+#     total_tokens = total_input + total_output
+    
+#     # Estimate cost based on model
+#     if st.session_state.model_name == "gpt-4o-mini":
+#         input_cost_per_1k = 0.0003  # $0.0003 per 1K input tokens
+#         output_cost_per_1k = 0.0006  # $$0.0006 per 1K output tokens
+#     elif "gpt-4" in st.session_state.model_name:
+#         input_cost_per_1k = 0.005 # $0.30 per 1K input tokens
+#         output_cost_per_1k = 0.60  # $0.60 per 1K output tokens
+#     else:  # Assume gpt-3.5-turbo pricing
+#         input_cost_per_1k = 0.0015  # $0.0015 per 1K input tokens
+#         output_cost_per_1k = 0.015   # $0.002 per 1K output tokens
+    
+#     estimated_cost = (total_input / 1000 * input_cost_per_1k) + (total_output / 1000 * output_cost_per_1k)
+    
+#     st.subheader("📊 Token Usage Statistics")
+    
+#     col1, col2, col3 = st.columns(3)
+    
+#     with col1:
+#         st.metric("Input Tokens", f"{total_input:,}")
+    
+#     with col2:
+#         st.metric("Output Tokens", f"{total_output:,}")
+    
+#     with col3:
+#         st.metric("Total Tokens", f"{total_tokens:,}")
+    
+#     st.markdown(f"**Estimated Cost:** ${estimated_cost:.4f}")
+    
+#     return total_tokens
+
+# # Function to parse and normalize tech stacks
+# def parse_tech_stack(stack):
+#     if pd.isna(stack) or stack == "" or stack is None:
+#         return set()
+#     if isinstance(stack, set):
+#         return stack
+#     try:
+#         # Handle potential string representation of sets
+#         if isinstance(stack, str) and stack.startswith("{") and stack.endswith("}"):
+#             # This could be a string representation of a set
+#             items = stack.strip("{}").split(",")
+#             return set(item.strip().strip("'\"") for item in items if item.strip())
+#         return set(map(lambda x: x.strip().lower(), str(stack).split(',')))
+#     except Exception as e:
+#         st.error(f"Error parsing tech stack: {e}")
+#         return set()
+
+# def display_tech_stack(stack_set):
+#     if isinstance(stack_set, set):
+#         return ", ".join(sorted(stack_set))
+#     return str(stack_set)
+
+# def get_matching_candidates(job_stack, candidates_df):
+#     """Find candidates with matching tech stack for a specific job"""
+#     matched = []
+#     job_stack_set = parse_tech_stack(job_stack)
+    
+#     for _, candidate in candidates_df.iterrows():
+#         candidate_stack = parse_tech_stack(candidate['Key Tech Stack'])
+#         common = job_stack_set & candidate_stack
+#         if len(common) >= 2:
+#             matched.append({
+#                 "Name": candidate["Full Name"],
+#                 "URL": candidate["LinkedIn URL"],
+#                 "Degree & Education": candidate["Degree & University"],
+#                 "Years of Experience": candidate["Years of Experience"],
+#                 "Current Title & Company": candidate['Current Title & Company'],
+#                 "Key Highlights": candidate["Key Highlights"],
+#                 "Location": candidate["Location (from most recent experience)"],
+#                 "Experience": str(candidate["Experience"]),
+#                 "Tech Stack": candidate_stack
+#             })
+#     return matched
+
+# def setup_llm():
+#     """Set up the LangChain LLM with structured output"""
+#     # Define the model to use
+#     model_name = "gpt-4o-mini"
+    
+#     # Store model name in session state for token calculation
+#     if 'model_name' not in st.session_state:
+#         st.session_state.model_name = model_name
+    
+#     # Create LLM instance
+#     llm = ChatOpenAI(
+#         model=model_name,
+#         temperature=0,
+#         max_tokens=None,
+#         timeout=None,
+#         max_retries=2,
+#     )
+    
+#     # Create structured output
+#     sum_llm = llm.with_structured_output(Shortlist)
+    
+#     # Create system prompt
+#     system = """You are an expert Recruitor, your task is to analyse the Candidate profile and determine if it matches with the job details and provide a score(out of 10) indicating how compatible the
+#     the profile is according to job.
+# Try to ensure following points while estimating the candidate's fit score:
+# For education:
+# Tier1 - MIT, Stanford, CMU, UC Berkeley, Caltech, Harvard, IIT Bombay, IIT Delhi, Princeton, UIUC, University of Washington, Columbia, University of Chicago, Cornell, University of Michigan (Ann Arbor), UT Austin - Maximum points
+# Tier2 - UC Davis, Georgia Tech, Purdue, UMass Amherst,etc - Moderate points
+# Tier3 - Unknown or unranked institutions - Lower points or reject
+
+# Startup Experience Requirement:
+# Candidates must have worked  as a direct employee at a VC-backed startup (Seed to series C/D)
+# preferred - Y Combinator, Sequoia,a16z,Accel,Founders Fund,LightSpeed,Greylock,Benchmark,Index Ventures,etc.   
+
+#     The fit score signifies based on following metrics:
+#     1–5 - Poor Fit - Auto-reject
+#     6–7 - Weak Fit - Auto-reject
+#     8.0–8.7 - Moderate Fit - Auto-reject
+#     8.8–10 - STRONG Fit - Include in results
+#     """
+    
+#     # Create query prompt
+#     query_prompt = ChatPromptTemplate.from_messages([
+#         ("system", system),
+#         ("human", """
+#     You are an expert Recruitor, your task is to determine if the user is a correct match for the given job or not.
+#           For this you will be provided with the follwing inputs of job and candidates:
+#     Job Details
+#     Company: {Company}
+#     Role: {Role}
+#     About Company: {desc}
+#     Locations: {Locations}
+#     Tech Stack: {Tech_Stack}
+#     Industry: {Industry}
+    
+
+#     Candidate Details:
+#     Full Name: {Full_Name}
+#     LinkedIn URL: {LinkedIn_URL}
+#     Current Title & Company: {Current_Title_Company}
+#     Years of Experience: {Years_of_Experience}
+#     Degree & University: {Degree_University}
+#     Key Tech Stack: {Key_Tech_Stack}
+#     Key Highlights: {Key_Highlights}
+#     Location (from most recent experience): {cand_Location}
+#     Past_Experience: {Experience}
+
+
+#     Answer in the structured manner as per the schema.
+#     If any parameter is Unknown try not to include in the summary, only include those parameters which are known.
+#     """),
+#     ])
+    
+#     # Chain the prompt and LLM
+#     cat_class = query_prompt | sum_llm
+    
+#     return cat_class
+
+# def call_llm(candidate_data, job_data, llm_chain):
+#     """Call the actual LLM to evaluate the candidate"""
+#     try:
+#         # Convert tech stacks to strings for the LLM payload
+#         job_tech_stack = job_data.get("Tech_Stack", set())
+#         candidate_tech_stack = candidate_data.get("Tech Stack", set())
+        
+#         if isinstance(job_tech_stack, set):
+#             job_tech_stack = ", ".join(sorted(job_tech_stack))
+            
+#         if isinstance(candidate_tech_stack, set):
+#             candidate_tech_stack = ", ".join(sorted(candidate_tech_stack))
+        
+#         # Prepare payload for LLM
+#         payload = {
+#             "Company": job_data.get("Company", ""),
+#             "Role": job_data.get("Role", ""),
+#             "desc": job_data.get("desc", ""),
+#             "Locations": job_data.get("Locations", ""),
+#             "Tech_Stack": job_tech_stack,
+#             "Industry": job_data.get("Industry", ""),
+
+#             "Full_Name": candidate_data.get("Name", ""),
+#             "LinkedIn_URL": candidate_data.get("URL", ""),
+#             "Current_Title_Company": candidate_data.get("Current Title & Company", ""),
+#             "Years_of_Experience": candidate_data.get("Years of Experience", ""),
+#             "Degree_University": candidate_data.get("Degree & Education", ""),
+#             "Key_Tech_Stack": candidate_tech_stack,
+#             "Key_Highlights": candidate_data.get("Key Highlights", ""),
+#             "cand_Location": candidate_data.get("Location", ""),
+#             "Experience": candidate_data.get("Experience", "")
+#         }
+        
+#         # Convert payload to a string for token calculation
+#         payload_str = json.dumps(payload)
+        
+#         # Calculate input tokens
+#         input_tokens = calculate_tokens(payload_str, st.session_state.model_name)
+        
+#         # Call LLM
+#         response = llm_chain.invoke(payload)
+#         print(candidate_data.get("Experience", ""))
+        
+#         # Convert response to string for token calculation
+#         response_str = f"""
+#         candidate_name: {response.candidate_name}
+#         candidate_url: {response.candidate_url}
+#         candidate_summary: {response.candidate_summary}
+#         candidate_location: {response.candidate_location}
+#         fit_score: {response.fit_score}
+#         justification: {response.justification}
+#         """
+        
+#         # Calculate output tokens
+#         output_tokens = calculate_tokens(response_str, st.session_state.model_name)
+        
+#         # Update token counts in session state
+#         if 'total_input_tokens' not in st.session_state:
+#             st.session_state.total_input_tokens = 0
+#         if 'total_output_tokens' not in st.session_state:
+#             st.session_state.total_output_tokens = 0
+        
+#         st.session_state.total_input_tokens += input_tokens
+#         st.session_state.total_output_tokens += output_tokens
+        
+#         # Return response in expected format
+#         return {
+#             "candidate_name": response.candidate_name,
+#             "candidate_url": response.candidate_url,
+#             "candidate_summary": response.candidate_summary,
+#             "candidate_location": response.candidate_location,
+#             "fit_score": response.fit_score,
+#             "justification": response.justification
+#         }
+#     except Exception as e:
+#         st.error(f"Error calling LLM: {e}")
+#         # Fallback to a default response
+#         return {
+#             "candidate_name": candidate_data.get("Name", "Unknown"),
+#             "candidate_url": candidate_data.get("URL", ""),
+#             "candidate_summary": "Error processing candidate profile",
+#             "candidate_location": candidate_data.get("Location", "Unknown"),
+#             "fit_score": 0.0,
+#             "justification": f"Error in LLM processing: {str(e)}"
+#         }
+
+# def process_candidates_for_job(job_row, candidates_df, llm_chain=None):
+#     """Process candidates for a specific job using the LLM"""
+#     # Reset token counters for this job
+#     st.session_state.total_input_tokens = 0
+#     st.session_state.total_output_tokens = 0
+    
+#     if llm_chain is None:
+#         with st.spinner("Setting up LLM..."):
+#             llm_chain = setup_llm()
+    
+#     selected_candidates = []
+    
+#     try:
+#         # Get job-specific data
+#         job_data = {
+#             "Company": job_row["Company"],
+#             "Role": job_row["Role"],
+#             "desc": job_row.get("One liner", ""),
+#             "Locations": job_row.get("Locations", ""),
+#             "Tech_Stack": job_row["Tech Stack"],
+#             "Industry": job_row.get("Industry", "")
+#         }
+        
+#         # Find matching candidates for this job
+#         with st.spinner("Finding matching candidates based on tech stack..."):
+#             matching_candidates = get_matching_candidates(job_row["Tech Stack"], candidates_df)
+        
+#         if not matching_candidates:
+#             st.warning("No candidates with matching tech stack found for this job.")
+#             return []
+        
+#         st.success(f"Found {len(matching_candidates)} candidates with matching tech stack.")
+        
+#         # Create progress elements
+#         candidates_progress = st.progress(0)
+#         candidate_status = st.empty()
+        
+#         # Process each candidate
+#         for i, candidate_data in enumerate(matching_candidates):
+#             # Update progress
+#             candidates_progress.progress((i + 1) / len(matching_candidates))
+#             candidate_status.text(f"Evaluating candidate {i+1}/{len(matching_candidates)}: {candidate_data.get('Name', 'Unknown')}")
+            
+#             # Process the candidate with the LLM
+#             response = call_llm(candidate_data, job_data, llm_chain)
+            
+#             response_dict = {
+#                 "Name": response["candidate_name"],
+#                 "LinkedIn": response["candidate_url"],
+#                 "summary": response["candidate_summary"],
+#                 "Location": response["candidate_location"],
+#                 "Fit Score": response["fit_score"],
+#                 "justification": response["justification"],
+#                 # Add back original candidate data for context
+#                 "Educational Background": candidate_data.get("Degree & Education", ""),
+#                 "Years of Experience": candidate_data.get("Years of Experience", ""),
+#                 "Current Title & Company": candidate_data.get("Current Title & Company", "")
+#             }
+            
+#             # Add to selected candidates if score is high enough
+#             if response["fit_score"] >= 8.8:
+#                 selected_candidates.append(response_dict)
+#                 st.markdown(response_dict)
+#             else:
+#                 st.write(f"Rejected candidate: {response_dict['Name']} with score: {response['fit_score']}")
+        
+#         # Clear progress indicators
+#         candidates_progress.empty()
+#         candidate_status.empty()
+        
+#         # Show results
+#         if selected_candidates:
+#             st.success(f"✅ Found {len(selected_candidates)} suitable candidates for this job!")
+#         else:
+#             st.info("No candidates met the minimum fit score threshold for this job.")
+        
+#         # Token usage is now displayed in display_job_selection when showing results
+#         return selected_candidates
+        
+#     except Exception as e:
+#         st.error(f"Error processing job: {e}")
+#         return []
+
+# def main():
+#     st.title("👨‍💻 Candidate Matching App")
+    
+#     # Initialize session state
+#     if 'processed_jobs' not in st.session_state:
+#         st.session_state.processed_jobs = {}
+    
+#     st.write("""
+#     This app matches job listings with candidate profiles based on tech stack and other criteria.
+#     Select a job to find matching candidates.
+#     """)
+    
+#     # API Key input
+#     with st.sidebar:
+#         st.header("API Configuration")
+#         api_key = st.text_input("Enter OpenAI API Key", type="password")
+#         if api_key:
+#             os.environ["OPENAI_API_KEY"] = api_key
+#             st.success("API Key set!")
+#         else:
+#             st.warning("Please enter OpenAI API Key to use LLM features")
+            
+#     # Show API key warning if not set
+#     SERVICE_ACCOUNT_FILE = 'src/synapse-recruitment-e94255ca76fd.json'
+#     SCOPES = ['https://www.googleapis.com/auth/spreadsheets']
+#     creds = service_account.Credentials.from_service_account_file(SERVICE_ACCOUNT_FILE, scopes=SCOPES)
+#     gc = gspread.authorize(creds)
+#     job_sheet = gc.open_by_key('1BZlvbtFyiQ9Pgr_lpepDJua1ZeVEqrCLjssNd6OiG9k')
+#     candidates_sheet = gc.open_by_key('1u_9o5f0MPHFUSScjEcnA8Lojm4Y9m9LuWhvjYm6ytF4')
+
+#     if not api_key:
+#         st.warning("⚠️ You need to provide an OpenAI API key in the sidebar to use this app.")
+
+#     if api_key:
+#         try:
+#             # Load data from Google Sheets
+#             job_worksheet = job_sheet.worksheet('paraform_jobs_formatted')
+#             job_data = job_worksheet.get_all_values()
+#             candidate_worksheet = candidates_sheet.worksheet('transformed_candidates_updated')
+#             candidate_data = candidate_worksheet.get_all_values()
+            
+#             # Convert to DataFrames
+#             jobs_df = pd.DataFrame(job_data[1:], columns=job_data[0])
+#             candidates_df = pd.DataFrame(candidate_data[1:], columns=candidate_data[0])
+#             candidates_df = candidates_df.fillna("Unknown")
+            
+#             # Display data preview
+#             with st.expander("Preview uploaded data"):
+#                 st.subheader("Jobs Data Preview")
+#                 st.dataframe(jobs_df.head(3))
+                
+#                 st.subheader("Candidates Data Preview")
+#                 st.dataframe(candidates_df.head(3))
+            
+#             # Map column names if needed
+#             column_mapping = {
+#                 "Full Name": "Full Name",
+#                 "LinkedIn URL": "LinkedIn URL",
+#                 "Current Title & Company": "Current Title & Company",
+#                 "Years of Experience": "Years of Experience",
+#                 "Degree & University": "Degree & University",
+#                 "Key Tech Stack": "Key Tech Stack", 
+#                 "Key Highlights": "Key Highlights",
+#                 "Location (from most recent experience)": "Location (from most recent experience)"
+#             }
+            
+#             # Rename columns if they don't match expected
+#             candidates_df = candidates_df.rename(columns={
+#                 col: mapping for col, mapping in column_mapping.items() 
+#                 if col in candidates_df.columns and col != mapping
+#             })
+            
+#             # Now, instead of processing all jobs upfront, we'll display job selection
+#             # and only process the selected job when the user chooses it
+#             display_job_selection(jobs_df, candidates_df)
+            
+#         except Exception as e:
+#             st.error(f"Error processing files: {e}")
+    
+#     st.divider()
+    
+
+# def display_job_selection(jobs_df, candidates_df):
+#     # Store the LLM chain as a session state to avoid recreating it
+#     if 'llm_chain' not in st.session_state:
+#         st.session_state.llm_chain = None
+    
+#     st.subheader("Select a job to view potential matches")
+    
+#     # Create job options - but don't compute matches yet
+#     job_options = []
+#     for i, row in jobs_df.iterrows():
+#         job_options.append(f"{row['Role']} at {row['Company']}")
+    
+#     if job_options:
+#         selected_job_index = st.selectbox("Jobs:", 
+#                                       range(len(job_options)),
+#                                       format_func=lambda x: job_options[x])
+        
+#         # Display job details
+#         job_row = jobs_df.iloc[selected_job_index]
+        
+#         # Parse tech stack for display
+#         job_row_stack = parse_tech_stack(job_row["Tech Stack"])
+        
+#         col1, col2 = st.columns([2, 1])
+        
+#         with col1:
+#             st.subheader(f"Job Details: {job_row['Role']}")
+            
+#             job_details = {
+#                 "Company": job_row["Company"],
+#                 "Role": job_row["Role"],
+#                 "Description": job_row.get("One liner", "N/A"),
+#                 "Locations": job_row.get("Locations", "N/A"),
+#                 "Industry": job_row.get("Industry", "N/A"),
+#                 "Tech Stack": display_tech_stack(job_row_stack)
+#             }
+            
+#             for key, value in job_details.items():
+#                 st.markdown(f"**{key}:** {value}")
+        
+#         # Create a key for this job in session state
+#         job_key = f"job_{selected_job_index}_processed"
+        
+#         if job_key not in st.session_state:
+#             st.session_state[job_key] = False
+        
+#         # Add a process button for this job
+#         if not st.session_state[job_key]:
+#             if st.button(f"Find Matching Candidates for this Job"):
+#                 if "OPENAI_API_KEY" not in os.environ or not os.environ["OPENAI_API_KEY"]:
+#                     st.error("Please enter your OpenAI API key in the sidebar before processing")
+#                 else:
+#                     # Process candidates for this job (only when requested)
+#                     selected_candidates = process_candidates_for_job(
+#                         job_row, 
+#                         candidates_df,
+#                         st.session_state.llm_chain
+#                     )
+                    
+#                     # Store the results and set as processed
+#                     if 'Selected_Candidates' not in st.session_state:
+#                         st.session_state.Selected_Candidates = {}
+#                     st.session_state.Selected_Candidates[selected_job_index] = selected_candidates
+#                     st.session_state[job_key] = True
+                    
+#                     # Store the LLM chain for reuse
+#                     if st.session_state.llm_chain is None:
+#                         st.session_state.llm_chain = setup_llm()
+                    
+#                     # Force refresh
+#                     st.rerun()
+        
+#         # Display selected candidates if already processed
+#         if st.session_state[job_key] and 'Selected_Candidates' in st.session_state:
+#             selected_candidates = st.session_state.Selected_Candidates.get(selected_job_index, [])
+            
+#             # Display selected candidates
+#             st.subheader("Selected Candidates")
+            
+#             # Display token usage statistics (will persist until job is changed)
+#             if 'total_input_tokens' in st.session_state and 'total_output_tokens' in st.session_state:
+#                 display_token_usage()
+            
+#             if len(selected_candidates) > 0:
+#                 for i, candidate in enumerate(selected_candidates):
+#                     with st.expander(f"{i+1}. {candidate['Name']} (Score: {candidate['Fit Score']})"):
+#                         col1, col2 = st.columns([3, 1])
+                        
+#                         with col1:
+#                             st.markdown(f"**Summary:** {candidate['summary']}")
+#                             st.markdown(f"**Current:** {candidate['Current Title & Company']}")
+#                             st.markdown(f"**Education:** {candidate['Educational Background']}")
+#                             st.markdown(f"**Experience:** {candidate['Years of Experience']}")
+#                             st.markdown(f"**Location:** {candidate['Location']}")
+#                             st.markdown(f"**[LinkedIn Profile]({candidate['LinkedIn']})**")
+                        
+#                         with col2:
+#                             st.markdown(f"**Fit Score:** {candidate['Fit Score']}")
+                        
+#                         st.markdown("**Justification:**")
+#                         st.info(candidate['justification'])
+#             else:
+#                 st.info("No candidates met the minimum score threshold (8.8) for this job.")
+                
+#                 # We don't show tech-matched candidates here since they are generated
+#                 # during the LLM matching process now
+            
+#             # Add a reset button to start over
+#             if st.button("Reset and Process Again"):
+#                 # Don't reset token counters here - we want them to persist
+#                 st.session_state[job_key] = False
+#                 st.rerun()
+
+# if __name__ == "__main__":
+#     main()
+
+
+
 import streamlit as st
 import pandas as pd
 import json
@@ -22,7 +604,7 @@ st.set_page_config(
 
 # Define pydantic model for structured output
 class Shortlist(BaseModel):
-    fit_score: float = Field(description="A score between 0 and 10 indicating how closely the candidate profile matches the job requirements.")
+    fit_score: float = Field(description="A score between 0 and 10 indicating how closely the candidate profile matches the job requirements upto 3 decimal points.")
     candidate_name: str = Field(description="The name of the candidate.")
     candidate_url: str = Field(description="The URL of the candidate's LinkedIn profile.")
     candidate_summary: str = Field(description="A brief summary of the candidate's skills and experience along with its educational background.")
@@ -145,7 +727,7 @@ def setup_llm():
     # Create LLM instance
     llm = ChatOpenAI(
         model=model_name,
-        temperature=0,
+        temperature=0.3,
         max_tokens=None,
         timeout=None,
         max_retries=2,
@@ -163,6 +745,7 @@ Tier1 - MIT, Stanford, CMU, UC Berkeley, Caltech, Harvard, IIT Bombay, IIT Delhi
 Tier2 - UC Davis, Georgia Tech, Purdue, UMass Amherst,etc - Moderate points
 Tier3 - Unknown or unranked institutions - Lower points or reject
 
+
 Startup Experience Requirement:
 Candidates must have worked  as a direct employee at a VC-backed startup (Seed to series C/D)
 preferred - Y Combinator, Sequoia,a16z,Accel,Founders Fund,LightSpeed,Greylock,Benchmark,Index Ventures,etc.   
@@ -172,13 +755,17 @@ preferred - Y Combinator, Sequoia,a16z,Accel,Founders Fund,LightSpeed,Greylock,B
     6–7 - Weak Fit - Auto-reject
     8.0–8.7 - Moderate Fit - Auto-reject
     8.8–10 - STRONG Fit - Include in results
+
+    Each candidate's fit score should be calculated based on a weighted evaluation of their background and **must be distinct even if candidates have similar profiles**. You may use slight variations to reflect nuanced differences.
     """
     
     # Create query prompt
     query_prompt = ChatPromptTemplate.from_messages([
         ("system", system),
         ("human", """
-    You are an expert Recruitor, your task is to determine if the user is a correct match for the given job or not.
+    You are an expert Recruitor. Your task is to determine if the candidate matches the given job. 
+Provide the score as a `float` rounded to exactly **three decimal places** (e.g., 8.943, 9.211, etc.). 
+Avoid rounding to whole or one-decimal numbers. Every candidate should have a **unique** fit score.
           For this you will be provided with the follwing inputs of job and candidates:
     Job Details
     Company: {Company}
@@ -203,6 +790,7 @@ preferred - Y Combinator, Sequoia,a16z,Accel,Founders Fund,LightSpeed,Greylock,B
 
     Answer in the structured manner as per the schema.
     If any parameter is Unknown try not to include in the summary, only include those parameters which are known.
+    The `fit_score` must be a float with **exactly three decimal digits** (e.g. 8.812, 9.006). Do not round to 1 or 2 decimals.
     """),
     ])
     
@@ -260,7 +848,7 @@ def call_llm(candidate_data, job_data, llm_chain):
         candidate_url: {response.candidate_url}
         candidate_summary: {response.candidate_summary}
         candidate_location: {response.candidate_location}
-        fit_score: {response.fit_score}
+        fit_score: {float(f"{response.fit_score:.3f}")}
         justification: {response.justification}
         """
         
@@ -348,7 +936,7 @@ def process_candidates_for_job(job_row, candidates_df, llm_chain=None):
                 "LinkedIn": response["candidate_url"],
                 "summary": response["candidate_summary"],
                 "Location": response["candidate_location"],
-                "Fit Score": response["fit_score"],
+                "Fit Score": float(f"{response['fit_score']:.3f}"),
                 "justification": response["justification"],
                 # Add back original candidate data for context
                 "Educational Background": candidate_data.get("Degree & Education", ""),
@@ -357,7 +945,7 @@ def process_candidates_for_job(job_row, candidates_df, llm_chain=None):
             }
             
             # Add to selected candidates if score is high enough
-            if response["fit_score"] >= 8.8:
+            if response["fit_score"] >= 8.800:
                 selected_candidates.append(response_dict)
                 st.markdown(response_dict)
             else:
@@ -454,7 +1042,7 @@ def main():
             
             # Now, instead of processing all jobs upfront, we'll display job selection
             # and only process the selected job when the user chooses it
-            display_job_selection(jobs_df, candidates_df)
+            display_job_selection(jobs_df, candidates_df, job_sheet)
             
         except Exception as e:
             st.error(f"Error processing files: {e}")
@@ -462,14 +1050,16 @@ def main():
     st.divider()
     
 
-def display_job_selection(jobs_df, candidates_df):
-    # Store the LLM chain as a session state to avoid recreating it
+def display_job_selection(jobs_df, candidates_df, sh):
+    # Initialize session state variables if they don't exist
+    if 'Selected_Candidates' not in st.session_state:
+        st.session_state.Selected_Candidates = {}
     if 'llm_chain' not in st.session_state:
-        st.session_state.llm_chain = None
+        st.session_state.llm_chain = setup_llm()
     
     st.subheader("Select a job to view potential matches")
     
-    # Create job options - but don't compute matches yet
+    # Create job options
     job_options = []
     for i, row in jobs_df.iterrows():
         job_options.append(f"{row['Role']} at {row['Company']}")
@@ -508,6 +1098,25 @@ def display_job_selection(jobs_df, candidates_df):
         if job_key not in st.session_state:
             st.session_state[job_key] = False
         
+        # Create worksheet name
+        sheet_name = f"{job_row['Role']} at {job_row['Company']}".strip()[:100]
+        
+        # Check if worksheet exists and has data
+        worksheet_exists = False
+        existing_candidates = []
+        
+        try:
+            cand_worksheet = sh.worksheet(sheet_name)
+            worksheet_exists = True
+            # Get existing data if worksheet exists
+            existing_data = cand_worksheet.get_all_values()
+            if len(existing_data) > 1:  # Has data beyond header
+                existing_candidates = existing_data[1:]
+                st.session_state[job_key] = True
+                # Don't show the info message about existing data
+        except gspread.exceptions.WorksheetNotFound:
+            pass
+        
         # Add a process button for this job
         if not st.session_state[job_key]:
             if st.button(f"Find Matching Candidates for this Job"):
@@ -515,65 +1124,95 @@ def display_job_selection(jobs_df, candidates_df):
                     st.error("Please enter your OpenAI API key in the sidebar before processing")
                 else:
                     # Process candidates for this job (only when requested)
-                    selected_candidates = process_candidates_for_job(
-                        job_row, 
-                        candidates_df,
-                        st.session_state.llm_chain
-                    )
+                    with st.spinner("Processing candidates..."):
+                        selected_candidates = process_candidates_for_job(
+                            job_row, 
+                            candidates_df,
+                            st.session_state.llm_chain
+                        )
+                        selected_candidates.sort(key=lambda x: x["Fit Score"], reverse=True)
                     
-                    # Store the results and set as processed
-                    if 'Selected_Candidates' not in st.session_state:
-                        st.session_state.Selected_Candidates = {}
-                    st.session_state.Selected_Candidates[selected_job_index] = selected_candidates
-                    st.session_state[job_key] = True
-                    
-                    # Store the LLM chain for reuse
-                    if st.session_state.llm_chain is None:
-                        st.session_state.llm_chain = setup_llm()
-                    
-                    # Force refresh
-                    st.rerun()
+                        # Only create worksheet if we have candidates
+                        if selected_candidates:
+                            try:
+                                if not worksheet_exists:
+                                    cand_worksheet = sh.add_worksheet(title=sheet_name, rows=10000, cols=50)
+                                
+                                # Prepare data for Google Sheet
+                                headers = list(selected_candidates[0].keys())
+                                rows = [headers] + [list(candidate.values()) for candidate in selected_candidates]
+                                
+                                # Clear existing data if any
+                                cand_worksheet.clear()
+                                
+                                # Write data to the worksheet
+                                cand_worksheet.update('A1', rows)
+                                
+                                st.success(f"Successfully processed {len(selected_candidates)} candidates")
+                            except Exception as e:
+                                st.error(f"Error writing to Google Sheet: {e}")
+                        
+                        # Store the results and set as processed
+                        st.session_state.Selected_Candidates[selected_job_index] = selected_candidates
+                        st.session_state[job_key] = True
+                        
+                        # Force refresh
+                        st.rerun()
         
         # Display selected candidates if already processed
-        if st.session_state[job_key] and 'Selected_Candidates' in st.session_state:
-            selected_candidates = st.session_state.Selected_Candidates.get(selected_job_index, [])
+        if st.session_state[job_key]:
+            if existing_candidates:
+                # Convert existing worksheet data to our format
+                headers = existing_data[0]
+                selected_candidates = []
+                for row in existing_data[1:]:
+                    candidate = dict(zip(headers, row))
+                    selected_candidates.append(candidate)
+                st.session_state.Selected_Candidates[selected_job_index] = selected_candidates
+            elif 'Selected_Candidates' in st.session_state:
+                selected_candidates = st.session_state.Selected_Candidates.get(selected_job_index, [])
+            else:
+                selected_candidates = []
             
             # Display selected candidates
             st.subheader("Selected Candidates")
             
-            # Display token usage statistics (will persist until job is changed)
-            if 'total_input_tokens' in st.session_state and 'total_output_tokens' in st.session_state:
+            # Display token usage statistics (only if we processed with LLM)
+            if not existing_candidates and 'total_input_tokens' in st.session_state and 'total_output_tokens' in st.session_state:
                 display_token_usage()
             
             if len(selected_candidates) > 0:
                 for i, candidate in enumerate(selected_candidates):
-                    with st.expander(f"{i+1}. {candidate['Name']} (Score: {candidate['Fit Score']})"):
+                    with st.expander(f"{i+1}. {candidate['Name']} (Score: {candidate.get('Fit Score', 'N/A')})"):
                         col1, col2 = st.columns([3, 1])
                         
                         with col1:
-                            st.markdown(f"**Summary:** {candidate['summary']}")
-                            st.markdown(f"**Current:** {candidate['Current Title & Company']}")
-                            st.markdown(f"**Education:** {candidate['Educational Background']}")
-                            st.markdown(f"**Experience:** {candidate['Years of Experience']}")
-                            st.markdown(f"**Location:** {candidate['Location']}")
-                            st.markdown(f"**[LinkedIn Profile]({candidate['LinkedIn']})**")
+                            st.markdown(f"**Summary:** {candidate.get('summary', 'N/A')}")
+                            st.markdown(f"**Current:** {candidate.get('Current Title & Company', 'N/A')}")
+                            st.markdown(f"**Education:** {candidate.get('Educational Background', 'N/A')}")
+                            st.markdown(f"**Experience:** {candidate.get('Years of Experience', 'N/A')}")
+                            st.markdown(f"**Location:** {candidate.get('Location', 'N/A')}")
+                            if 'LinkedIn' in candidate:
+                                st.markdown(f"**[LinkedIn Profile]({candidate['LinkedIn']})**")
                         
                         with col2:
-                            st.markdown(f"**Fit Score:** {candidate['Fit Score']}")
+                            if 'Fit Score' in candidate:
+                                st.markdown(f"**Fit Score:** {candidate['Fit Score']}")
                         
-                        st.markdown("**Justification:**")
-                        st.info(candidate['justification'])
+                        if 'justification' in candidate:
+                            st.markdown("**Justification:**")
+                            st.info(candidate['justification'])
             else:
-                st.info("No candidates met the minimum score threshold (8.8) for this job.")
-                
-                # We don't show tech-matched candidates here since they are generated
-                # during the LLM matching process now
+                st.info("No candidates found for this job.")
             
             # Add a reset button to start over
             if st.button("Reset and Process Again"):
-                # Don't reset token counters here - we want them to persist
+                # Reset this job's processing state
                 st.session_state[job_key] = False
+                if 'Selected_Candidates' in st.session_state and selected_job_index in st.session_state.Selected_Candidates:
+                    del st.session_state.Selected_Candidates[selected_job_index]
                 st.rerun()
 
+
 if __name__ == "__main__":
     main()