Update retriever.py

retriever.py

@@ -1,110 +1,38 @@
 import pandas as pd
-import gradio as gr
-from retriever import get_relevant_passages
-from reranker import rerank
+from sentence_transformers import SentenceTransformer, util
 
-# Load and clean CSV
-def clean_df(df):
-    df = df.copy()
-    
-    # Ensure clean URLs
-    # Check if the second column contains URLs or just IDs
-    second_col = df.iloc[:, 1].astype(str)
-    if second_col.str.contains('http').any() or second_col.str.contains('www').any():
-        df["url"] = second_col  # Already has full URLs
-    else:
-        # Create full URLs from IDs
-        df["url"] = "https://www.shl.com/" + second_col.str.replace(r'^[\s/]*', '', regex=True)
-    
-    df["remote_support"] = df.iloc[:, 2].map(lambda x: "Yes" if x == "T" else "No")
-    df["adaptive_support"] = df.iloc[:, 3].map(lambda x: "Yes" if x == "T" else "No")
+model = SentenceTransformer("all-MiniLM-L6-v2")
+
+def get_relevant_passages(query, df, top_k=20):
+    # Create a copy to avoid modifying the original dataframe
+    df_copy = df.copy()
     
-    # Handle test_type with error checking
-    df["test_type"] = df.iloc[:, 4].astype(str).str.split("\\n")
+    # Ensure URL field is properly formatted
+    if 'url' in df_copy.columns:
+        # Clean up URLs if needed
+        df_copy['url'] = df_copy['url'].astype(str)
+        # Ensure URLs start with http or https
+        mask = ~df_copy['url'].str.startswith(('http://', 'https://'))
+        df_copy.loc[mask, 'url'] = 'https://www.shl.com/' + df_copy.loc[mask, 'url'].str.lstrip('/')
     
-    df["description"] = df.iloc[:, 5]
+    # Format test_type for better representation
+    def format_test_type(test_types):
+        if isinstance(test_types, list):
+            return ', '.join(test_types)
+        return str(test_types)
     
-    # Extract duration with error handling
-    df["duration"] = pd.to_numeric(
-        df.iloc[:, 8].astype(str).str.extract(r'(\d+)')[0], 
-        errors='coerce'
-    )
+    # Concatenate all fields into a single string per row
+    corpus = df_copy.apply(
+        lambda row: f"{row['description']} "
+                   f"Test types: {format_test_type(row['test_type'])}. "
+                   f"Adaptive support: {row['adaptive_support']}. "
+                   f"Remote support: {row['remote_support']}. "
+                   f"Duration: {row['duration'] if pd.notna(row['duration']) else 'N/A'} minutes.",
+        axis=1
+    ).tolist()
     
-    return df[["url", "adaptive_support", "remote_support", "description", "duration", "test_type"]]
-
-try:
-    df = pd.read_csv("assesments.csv")
-    df_clean = clean_df(df)
-except Exception as e:
-    print(f"Error loading or cleaning data: {e}")
-    # Create an empty DataFrame with required columns as fallback
-    df_clean = pd.DataFrame(columns=["url", "adaptive_support", "remote_support", 
-                                     "description", "duration", "test_type"])
-
-def validate_and_fix_urls(candidates):
-    """Validates and fixes URLs in candidate assessments."""
-    for candidate in candidates:
-        # Ensure URL exists
-        if 'url' not in candidate or not candidate['url']:
-            candidate['url'] = 'https://www.shl.com/missing-url'
-            continue
-            
-        url = str(candidate['url'])
-        
-        # Fix URLs that are just numbers
-        if url.isdigit() or (url.startswith('https://www.shl.com') and url[len('https://www.shl.com'):].isdigit()):
-            candidate['url'] = f"https://www.shl.com/{url.replace('https://www.shl.com', '')}"
-            continue
-            
-        # Add protocol if missing
-        if not url.startswith(('http://', 'https://')):
-            candidate['url'] = f"https://{url}"
-            
-    return candidates
-
-def recommend(query):
-    if not query.strip():
-        return {"error": "Please enter a job description"}
+    corpus_embeddings = model.encode(corpus, convert_to_tensor=True)
+    query_embedding = model.encode(query, convert_to_tensor=True)
+    hits = util.semantic_search(query_embedding, corpus_embeddings, top_k=top_k)[0]
     
-    try:
-        # Print some debug info
-        print(f"Processing query: {query[:50]}...")
-        
-        top_k_df = get_relevant_passages(query, df_clean, top_k=20)
-        
-        # Debug: Check URLs in retrieved data
-        print(f"Retrieved {len(top_k_df)} assessments")
-        if not top_k_df.empty:
-            print(f"Sample URLs from retrieval: {top_k_df['url'].iloc[:3].tolist()}")
-        
-        candidates = top_k_df.to_dict(orient="records")
-        
-        # Additional URL validation before sending to reranker
-        for c in candidates:
-            if 'url' in c:
-                if not str(c['url']).startswith(('http://', 'https://')):
-                    c['url'] = f"https://www.shl.com/{str(c['url']).lstrip('/')}"
-        
-        result = rerank(query, candidates)
-        
-        # Post-process result to ensure URLs are properly formatted
-        if 'recommended_assessments' in result:
-            result['recommended_assessments'] = validate_and_fix_urls(result['recommended_assessments'])
-            
-        return result
-    except Exception as e:
-        import traceback
-        error_details = traceback.format_exc()
-        print(f"Error: {str(e)}\n{error_details}")
-        return {"error": f"Error processing request: {str(e)}"}
-
-iface = gr.Interface(
-    fn=recommend,
-    inputs=gr.Textbox(label="Enter Job Description", lines=4),
-    outputs="json",
-    title="SHL Assessment Recommender",
-    description="Paste a job description to get the most relevant SHL assessments."
-)
-
-if __name__ == "__main__":
-    iface.launch()
+    return df_copy.iloc[[hit['corpus_id'] for hit in hits]]