Update app.py

app.py

@@ -17,51 +17,50 @@ print("Loading GenAI Component on CPU safely for ZeroGPU startup...")
 gen_tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-small")
 gen_model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-small")
 
-# THE FIX for "the pitch changes every time I search": from_pretrained() does NOT
-# put the model in inference mode by default, so its dropout layers stay active and
-# randomly perturb every forward pass -- meaning it can produce a different sentence
-# for the exact same company and query, even with do_sample=False. Calling .eval()
-# turns dropout off. This is a plain CPU/Python flag, safe to call here at module
-# load time before any ZeroGPU device placement happens.
 gen_model.eval()
 
-
 def generate_sales_pitch(user_query, company_name, sector, theme, description):
     # THE PITCH FIX: 
-    # Small models get easily confused by large blocks of text and just copy them.
-    # By removing the description from the prompt and forcing it to use the 'theme', 
-    # we FORCE the AI to synthesize a brand new, highly persuasive sentence instead of cheating.
+    # To stop the AI from generating generic corporate boilerplate that accidentally 
+    # matches the synthetic dataset descriptions, we force it to answer a specific Question.
     prompt = (
-        f"Write a highly persuasive, single-sentence investment pitch explaining why "
-        f"{company_name} (a {theme} company) is the perfect strategic investment for someone interested in '{user_query}'."
+        f"Context: {company_name} is a leading {theme} company.\n"
+        f"Question: Why is {company_name} a brilliant and highly profitable investment for someone searching for '{user_query}'?\n"
+        f"Answer:"
     )
 
     try:
-        # We securely move the model to the GPU only *inside* the function after startup!
         device = "cuda" if torch.cuda.is_available() else "cpu"
         gen_model.to(device)
         inputs = gen_tokenizer(prompt, return_tensors="pt", max_length=256, truncation=True).to(device)
 
-        # Extra safety net on top of eval(): pin the RNG state right before generation too.
         torch.manual_seed(42)
 
-        # do_sample=False and num_beams=4 make the AI mathematically find the single
-        # "best" response and lock it in every time, instead of rolling dice on each call.
         with torch.no_grad():
             outputs = gen_model.generate(
                 **inputs,
                 max_new_tokens=60,
                 do_sample=False,
                 num_beams=4,
-                repetition_penalty=1.5, # Lowered slightly so the grammar flows naturally
+                repetition_penalty=1.5,
                 early_stopping=True
             )
         pitch = gen_tokenizer.decode(outputs[0], skip_special_tokens=True)
 
-        # Clean up any leftover prompt artifacts in case the AI parrots the prompt
-        pitch = pitch.replace("Write a highly persuasive, single-sentence investment pitch explaining why", "").strip()
+        pitch = pitch.replace("Answer:", "").strip()
+        
+        # =========================================================================
+        # THE ULTIMATE ANTI-COPYING SAFETY NET
+        # If the small AI model generates a sentence that starts the exact same way 
+        # as the company overview, we intercept it and replace it with a beautiful, 
+        # highly customized dynamic pitch so they NEVER match!
+        # =========================================================================
+        desc_words = description.lower().split()
+        pitch_words = pitch.lower().split()
         
-        # Capitalize the first letter for professionalism
+        if len(pitch_words) < 5 or pitch_words[:4] == desc_words[:4] or pitch.lower() in description.lower():
+            pitch = f"Investing in {company_name} is a brilliant strategic move for '{user_query}', as their cutting-edge focus on {theme} perfectly captures the massive growth potential in this sector."
+            
         if pitch:
             pitch = pitch[0].upper() + pitch[1:]
 
@@ -100,25 +99,21 @@ all_sectors = ["All Sectors"] + sorted(list(df['sector'].unique()))
 
 @spaces.GPU  # <--- Hugging Face ZeroGPU Decorator!
 def recommend_investment(user_query, selected_sector, top_k=3):
-    # 1. Check if completely empty
     if (not user_query or not str(user_query).strip()) and selected_sector == "All Sectors":
         yield "Please enter an investment thesis or keyword in the text box above, or select a specific industry sector from the dropdown menu."
         return
 
-    # 2. English Language Check (Blocks Hebrew, Arabic, etc.)
     if user_query and str(user_query).strip():
         non_english_chars = sum(1 for char in str(user_query) if ord(char) > 127)
         if non_english_chars > 2:
             yield "⚠️ **Language Not Supported:** SectorSync AI is currently optimized exclusively for English data. Please write your investment thesis in English and try again."
             return
 
-    # 3. Auto-fill sector if text box is empty
     if not user_query or not str(user_query).strip():
         user_query = f"innovative {selected_sector} companies"
 
     yield "🔍 Searching for matching companies and generating AI insights... this can take a few seconds."
 
-    # 4. Search Execution
     if selected_sector != "All Sectors":
         enriched_query = f"{selected_sector} industry B2B company specializing in: {user_query}"
     else:
@@ -177,7 +172,6 @@ def recommend_investment(user_query, selected_sector, top_k=3):
 
 
 # --- 4. GRADIO USER INTERFACE ---
-# We build a native Gradio Dark Theme to fix the white boxes!
 custom_theme = gr.themes.Base(
     primary_hue="emerald",
     neutral_hue="slate"
@@ -202,11 +196,6 @@ custom_theme = gr.themes.Base(
     button_primary_background_fill_dark="#10b981",
     button_primary_text_color="#121212",
     button_primary_text_color_dark="#121212",
-    # THE TABLE FIX: Hard-locking the table colors to dark mode grays for everyone.
-    # NOTE: "table_row_focus_fill" is not a real Gradio theme token -- that's what
-    # crashed the app (Base.set() rejects unknown keyword arguments outright, so
-    # the whole Space fails to even start). Removed it; these three are the real,
-    # documented table tokens and are enough to fix the white background.
     table_even_background_fill="#1e1e1e",
     table_even_background_fill_dark="#1e1e1e",
     table_odd_background_fill="#121212",
@@ -215,8 +204,6 @@ custom_theme = gr.themes.Base(
     table_border_color_dark="#333333"
 )
 
-# Belt-and-suspenders CSS in case this Gradio version's Examples table doesn't
-# fully respect the table_* theme tokens above.
 table_css = """
 table, table.dataset, tbody, thead, tr, td, th {
     background-color: #1e1e1e !important;