Update app.py

app.py

@@ -2,23 +2,20 @@ import gradio as gr
 import pandas as pd
 import json
 import os
-import requests
 import re 
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.metrics.pairwise import cosine_similarity
 from functools import lru_cache
 
 # ----------------------------
 # CONFIG
 # ----------------------------
 JSON_FILE = "form-submissions-1.json"
-MODEL_ID = "google/flan-t5-small" 
-# NOTE: HF_API_TOKEN MUST be set in your environment variables/Space secrets.
-HF_API_TOKEN = os.environ.get("HF_API_TOKEN") 
 FILTERED_CSV = "/tmp/filtered_candidates.csv"
-OUTPUT_FILE = "/tmp/outputs.csv"
-BATCH_SIZE = 50 
 
-if not HF_API_TOKEN:
-    pass 
+# Note: The following variables are no longer needed, as we are not using the Hugging Face API
+# MODEL_ID = "google/flan-t5-small" 
+# HF_API_TOKEN = os.environ.get("HF_API_TOKEN") 
 
 CATEGORIES = {
     "AI": [
@@ -35,58 +32,38 @@ CATEGORIES = {
 }
 
 # ----------------------------
-# LLM Call for Scoring (Focus: Role Experience ONLY)
+# Similarity Matching Function (REPLACING LLM)
 # ----------------------------
-@lru_cache(maxsize=512) 
-def score_candidate(candidate_str, category_name, job_titles_tuple):
-    if not HF_API_TOKEN:
-         print("API Token is missing. Returning score 0.")
-         return 0
-         
-    prompt = f"""
-You are an HR assistant. Your task is to rate a candidate's suitability based ONLY on their previous job roles.
-Rate the suitability of the following candidate on a scale of 1 (Lowest) to 10 (Highest).
-The score must reflect how closely the candidate's 'Roles' align with the target job titles.
-
-The target roles for the '{category_name}' category are: {list(job_titles_tuple)}
-
-Candidate JSON: {candidate_str}
-
-**Task**: Respond ONLY with the rating number (an integer from 1 to 10).
-"""
-    headers = {"Authorization": f"Bearer {HF_API_TOKEN}", "Content-Type": "application/json"}
+@lru_cache(maxsize=1) 
+def calculate_similarity_scores(df_candidates, category_name):
+    """
+    Calculates the cosine similarity between candidate roles and target job titles.
+    """
+    if df_candidates.empty:
+        return pd.Series([], dtype='float64')
+
+    # 1. Define the document corpus
+    target_roles = " ".join(CATEGORIES[category_name])
+    candidate_roles = df_candidates['Roles'].tolist()
     
-    payload = {
-        "inputs": prompt,
-        "parameters": {
-            "max_new_tokens": 5, 
-            "return_full_text": False,
-            "temperature": 0.1 
-        }
-    }
+    # 2. Create the corpus for vectorization
+    corpus = [target_roles] + candidate_roles
     
-    try:
-        response = requests.post(
-            f"https://api-inference.huggingface.co/models/{MODEL_ID}",
-            headers=headers,
-            data=json.dumps(payload),
-            timeout=60 
-        )
-        response.raise_for_status()
-        result = response.json()
-             
-        generated_text = result[0].get("generated_text", "0").strip()
-        
-        match = re.search(r'\d+', generated_text)
-        if match:
-            score = int(match.group(0))
-            return max(1, min(10, score)) 
-        
-        return 0 
-            
-    except Exception as e:
-        print(f"LLM scoring call failed for candidate (API/Network Error): {e}")
-        return 0 
+    # 3. Vectorize using TF-IDF (converts text to numerical features)
+    # Uses unigrams and bigrams for better matching (e.g., 'Data Scientist' vs 'Data' 'Scientist')
+    vectorizer = TfidfVectorizer(ngram_range=(1, 2), stop_words='english')
+    tfidf_matrix = vectorizer.fit_transform(corpus)
+    
+    # 4. Extract the vector for the target roles (the first row)
+    target_vector = tfidf_matrix[0]
+    candidate_vectors = tfidf_matrix[1:]
+    
+    # 5. Calculate Cosine Similarity between the target and all candidates
+    # The result is a matrix where [0, i] is the similarity score for candidate i
+    similarity_scores = cosine_similarity(target_vector, candidate_vectors).flatten()
+    
+    # Return scores as a Pandas Series aligned with the DataFrame index
+    return pd.Series(similarity_scores, index=df_candidates.index)
 
 # ----------------------------
 # Step 1: Filter by roles (Unchanged)
@@ -97,7 +74,7 @@ def filter_by_roles(category_name):
         with open(JSON_FILE, encoding="utf-8") as f:
             data = json.load(f)
     except FileNotFoundError:
-        return pd.DataFrame(), f"Error: JSON file '{JSON_FILE}' not found. The LLM can't proceed."
+        return pd.DataFrame(), f"Error: JSON file '{JSON_FILE}' not found. The application can't proceed."
         
     filtered = []
 
@@ -125,19 +102,17 @@ def filter_by_roles(category_name):
             })
             
     if not filtered:
-        return pd.DataFrame(), f"No candidates found matching roles for category '{category_name}'. The LLM can't proceed."
+        return pd.DataFrame(), f"No candidates found matching roles for category '{category_name}'. The application can't proceed."
 
     df = pd.DataFrame(filtered)
     df.to_csv(FILTERED_CSV, index=False)
-    return df, f"{len(df)} candidates filtered by role for category '{category_name}'. Ready for LLM scoring."
+    return df, f"{len(df)} candidates filtered by role for category '{category_name}'. Ready for Similarity Ranking."
 
 
 # ----------------------------
-# Step 2: LLM recommendations (Scoring, Sorting, and Output)
+# Step 2: Recommendations (Using Similarity Matching)
 # ----------------------------
-def llm_recommendations(category_name):
-    job_titles = CATEGORIES[category_name]
-
+def similarity_recommendations(category_name):
     if not os.path.exists(FILTERED_CSV):
         df_filtered, msg = filter_by_roles(category_name)
         if df_filtered.empty:
@@ -149,32 +124,16 @@ def llm_recommendations(category_name):
     if df_filtered.empty:
         return f"No filtered candidates found for category '{category_name}'. Run Step 1 first."
 
-    # Prepare for scoring
-    df_filtered_clean = df_filtered.fillna('N/A')
-    filtered_candidates = df_filtered_clean.to_dict(orient="records")
-    
-    scores = []
-    
-    for person in filtered_candidates:
-        candidate_info = {
-            "Name": person.get("Name"),
-            "Roles": person.get("Roles"),
-            "Skills": person.get("Skills") 
-        }
-        candidate_str = json.dumps(candidate_info)
-        
-        score = score_candidate(candidate_str, category_name, tuple(job_titles))
-        scores.append(score)
-
-    df_filtered["LLM_Score"] = scores
+    # --- CORE CHANGE: Calculate Similarity Scores ---
+    df_filtered["Similarity_Score"] = calculate_similarity_scores(df_filtered, category_name)
     
-    df_recommended = df_filtered[df_filtered["LLM_Score"] > 0].copy()
+    # Filter out candidates with zero relevance (can happen if roles are too generic)
+    df_recommended = df_filtered[df_filtered["Similarity_Score"] > 0].copy()
 
     if df_recommended.empty:
-        if not HF_API_TOKEN:
-            return "❌ LLM failed: The HF_API_TOKEN is not set or is invalid. Set the token and try again."
-        return f"LLM scored all candidates 0. The candidates' roles are deemed irrelevant by the LLM for '{category_name}'."
+        return f"All candidates had zero text similarity to the target roles for '{category_name}'. The current roles are not specific enough to match."
     
+    # Define salary parsing for tie-breaker
     def parse_salary(s):
         try:
             return float(str(s).replace("$","").replace(",","").replace("N/A", str(float('inf'))))
@@ -183,8 +142,9 @@ def llm_recommendations(category_name):
             
     df_recommended["Salary_sort"] = df_recommended["Salary"].apply(parse_salary)
 
+    # Sort: 1. Highest Similarity Score (descending), 2. Lowest Salary (ascending)
     df_top5 = df_recommended.sort_values(
-        by=['LLM_Score', 'Salary_sort'],
+        by=['Similarity_Score', 'Salary_sort'],
         ascending=[False, True] 
     ).head(5)
 
@@ -193,10 +153,12 @@ def llm_recommendations(category_name):
     output_text = f"Top {len(final_names)} Recommended Candidates for the '{category_name}' Category:\n\n"
     
     for i, name in enumerate(final_names):
-        score = df_top5.iloc[i]['LLM_Score']
-        output_text += f"{i+1}. {name} (Suitability Score: {score}/10)\n"
+        score = df_top5.iloc[i]['Similarity_Score']
+        # Display the score as a percentage for readability
+        score_percent = f"{score * 100:.2f}%"
+        output_text += f"{i+1}. {name} (Role Match: {score_percent})\n"
     
-    output_text += "\nThese candidates were ranked by the LLM based **only on the alignment of their previous job roles** with the target roles, using expected salary as a tie-breaker."
+    output_text += "\nThese candidates were ranked objectively based on the **text similarity** of their previous job roles to the target roles, using expected salary as a tie-breaker."
     
     return output_text
 
@@ -217,8 +179,8 @@ def show_first_candidates():
 # Gradio interface (Updated Heading and Launch)
 # ----------------------------
 with gr.Blocks() as app:
-    # 🚩 CHANGE: Updated Heading
-    gr.Markdown("# 🤖 Candidate Selection (Role-Based Scoring)") 
+    gr.Markdown("# 🏆 Candidate Selection (Objective Similarity Matching)") 
+    gr.Markdown("### **Reliable ranking using text similarity (TF-IDF & Cosine Similarity) - No LLM API required.**")
     
     gr.Markdown("#### 🔍 Raw JSON Preview: First 5 Candidates")
     gr.Dataframe(show_first_candidates(), label="First 5 JSON Entries")
@@ -234,11 +196,11 @@ with gr.Blocks() as app:
 
     gr.Markdown("---")
 
-    # Step 2: LLM Recommendations
-    llm_button = gr.Button("3. Get LLM Recommendations (Role Experience Ranking)")
-    llm_output_text = gr.Textbox(label="Top Candidate Recommendations Summary", lines=10, placeholder="Click 'Get LLM Recommendations' after Step 2 completes.")
-    llm_button.click(llm_recommendations, inputs=[category_dropdown], outputs=[llm_output_text])
+    # Step 2: Recommendations
+    # Changed function name to reflect the new logic
+    recommend_button = gr.Button("3. Rank Candidates by Role Similarity")
+    recommend_output_text = gr.Textbox(label="Top Candidate Recommendations Summary", lines=10, placeholder="Click 'Rank Candidates by Role Similarity' after Step 2 completes.")
+    recommend_button.click(similarity_recommendations, inputs=[category_dropdown], outputs=[recommend_output_text])
 
 if __name__ == "__main__":
-    # 🚩 CHANGE: Set share=True to generate a public link
     app.launch(share=True)