Update app.py

app.py

@@ -13,11 +13,20 @@ import spaces  # <--- Import the Hugging Face spaces library for Free GPU
 # --- 1. LOAD GENERATIVE AI COMPONENT ---
 device = "cuda" if torch.cuda.is_available() else "cpu"
 print(f"Loading GenAI Component on {device}...")
-
 # Using the smarter base model for high-quality text generation
 gen_tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-base")
 gen_model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-base").to(device)
 
+# 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
+# the model can genuinely produce a different sentence for the exact same company
+# and query, even with do_sample=False. Calling .eval() turns dropout off, which
+# combined with do_sample=False/num_beams below makes generation fully deterministic:
+# same company + same query -> same pitch, every single time.
+gen_model.eval()
+
+
 def generate_sales_pitch(user_query, company_name, sector, theme, description):
     # Feeding the company's description to the AI so it knows exactly what it's selling
     prompt = (
@@ -25,26 +34,29 @@ def generate_sales_pitch(user_query, company_name, sector, theme, description):
         f"Write one single, highly persuasive sentence explaining why investing in {company_name} "
         f"is the perfect choice for someone looking for '{user_query}'."
     )
-
     try:
         inputs = gen_tokenizer(prompt, return_tensors="pt", max_length=256, truncation=True).to(device)
-
-        # THE FIX: do_sample=False and num_beams=4 ensures the AI mathematically finds 
-        # the single "best" response and locks it in every time without changing.
-        outputs = gen_model.generate(
-            **inputs,
-            max_new_tokens=60,
-            do_sample=False,
-            num_beams=4,
-            repetition_penalty=2.0,
-            early_stopping=True
-        )
+        # Extra safety net: pin the RNG state right before generation too, in case any
+        # part of the model still touches it. With eval() already set this is now
+        # belt-and-suspenders, not the primary fix -- but it costs nothing to keep.
+        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=2.0,
+                early_stopping=True
+            )
         pitch = gen_tokenizer.decode(outputs[0], skip_special_tokens=True)
-        
+
         # Clean up any leftover prompt artifacts
         pitch = pitch.replace("Based on this company description:", "").strip()
         return pitch
-        
+
     except Exception:
         return f"{company_name} is an exceptional strategic match for '{user_query}' within the {sector} space."
 
@@ -66,15 +78,15 @@ if 'faiss_index' not in globals():
         faiss_index = faiss.IndexFlatL2(embeddings.shape[1])
         faiss_index.add(np.array(embeddings).astype('float32'))
     except FileNotFoundError:
-        print("Saved embeddings not scratch...")
+        print("Saved embeddings not found, encoding from scratch...")
         embeddings = embedding_model.encode(df['full_text'].tolist(), show_progress_bar=False)
         faiss_index = faiss.IndexFlatL2(embeddings.shape[1])
         faiss_index.add(np.array(embeddings).astype('float32'))
 
-
 # --- 3. RECOMMENDATION ENGINE LOGIC ---
 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
@@ -86,7 +98,7 @@ def recommend_investment(user_query, selected_sector, top_k=3):
     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."
+            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
@@ -122,9 +134,9 @@ def recommend_investment(user_query, selected_sector, top_k=3):
             sim_score = 1 / (1 + distances[0][i])
             matches.append((df.iloc[idx], sim_score))
 
-    output_markdown = f"### Top {len(matches)} AI-Recommended Matches for: *'{user_query}'*\n\n"
+    output_markdown = f"### Top {len(matches)} AI-Recommended Matches for: '{user_query}'\n\n"
     if selected_sector != "All Sectors":
-        output_markdown += f"**Filtered by Sector:** `{selected_sector}`\n\n---\n\n"
+        output_markdown += f"*Filtered by Sector:* ⁠ {selected_sector} ⁠\n\n---\n\n"
     else:
         output_markdown += "---\n\n"
 
@@ -145,18 +157,18 @@ def recommend_investment(user_query, selected_sector, top_k=3):
         tier_label = match_labels[rank - 1] if rank <= len(match_labels) else "Match"
 
         # The Fix: No bullet points, reads like a clean, professional report!
-        output_markdown += f"#### #{rank}. {c_name} (`{ticker}`) — *{sector}*\n\n"
-        output_markdown += f"**{tier_label}** (Similarity Score: `{score*100:.1f}%`)\n\n"
-        output_markdown += f"**Industry:** {theme}\n\n"
-        output_markdown += f"**GenAI Investment Pitch:** *'{pitch}'*\n\n"
-        output_markdown += f"**Company Overview:** {desc[:200]}...\n\n---\n\n"
+        output_markdown += f"#### #{rank}. {c_name} (⁠ {ticker} ⁠) — {sector}\n\n"
+        output_markdown += f"*{tier_label}* (Similarity Score: ⁠ {score*100:.1f}% ⁠)\n\n"
+        output_markdown += f"*Industry:* {theme}\n\n"
+        output_markdown += f"*GenAI Investment Pitch:* '{pitch}'\n\n"
+        output_markdown += f"*Company Overview:* {desc[:200]}...\n\n---\n\n"
 
         # Streams the results instantly!
         yield output_markdown
 
 
 # --- 4. GRADIO USER INTERFACE ---
-# Instead of CSS hacks, we build a native Gradio Dark Theme to fix the white boxes!
+# Native Gradio Dark Theme to fix the white boxes.
 custom_theme = gr.themes.Base(
     primary_hue="emerald",
     neutral_hue="slate"
@@ -177,16 +189,40 @@ custom_theme = gr.themes.Base(
     block_label_text_color_dark="#f0f4f8",
     input_background_fill="#2a2a2a",
     input_background_fill_dark="#2a2a2a",
+    input_border_color="#333333",
+    input_border_color_dark="#333333",
+    input_placeholder_color="#9ca3af",
+    input_placeholder_color_dark="#9ca3af",
     button_primary_background_fill="#10b981",
     button_primary_background_fill_dark="#10b981",
     button_primary_text_color="#121212",
-    button_primary_text_color_dark="#121212"
+    button_primary_text_color_dark="#121212",
+    # THE FIX for "the table is still white": the Quick Starters examples list
+    # renders as a table, and the theme's generic background tokens above don't
+    # touch it -- it has its own dedicated table_* tokens that must be set separately.
+    table_even_background_fill="#1a1a1a",
+    table_even_background_fill_dark="#1a1a1a",
+    table_odd_background_fill="#242424",
+    table_odd_background_fill_dark="#242424",
+    table_border_color="#333333",
+    table_border_color_dark="#333333",
 )
 
-with gr.Blocks(theme=custom_theme, title="SectorSync AI") as demo:
+# Belt-and-suspenders CSS in case this Gradio version's Examples table doesn't
+# fully respect the table_* theme tokens above -- forces its background/text
+# directly regardless of theme token coverage.
+table_css = """
+table, table.dataset, tbody, thead, tr, td, th {
+    background-color: #1a1a1a !important;
+    color: #f0f4f8 !important;
+    border-color: #333333 !important;
+}
+"""
+
+with gr.Blocks(theme=custom_theme, css=table_css, title="SectorSync AI") as demo:
     gr.Markdown("# SectorSync AI")
     gr.Markdown("Cut Through the Market Noise — Find Your Next Winning Stock in Seconds. Discover high-growth companies matching your investment thesis using FAISS Vector Search and Generative AI.")
-    gr.Markdown("*Similarity Score reflects how closely a company profile matches your query in AI-embedding space — a relative ranking signal, not a calibrated financial confidence rating.*")
+    gr.Markdown("Similarity Score reflects how closely a company profile matches your query in AI-embedding space — a relative ranking signal, not a calibrated financial confidence rating.")
 
     with gr.Row():
         with gr.Column(scale=2):
@@ -219,6 +255,7 @@ with gr.Blocks(theme=custom_theme, title="SectorSync AI") as demo:
     search_button.click(fn=recommend_investment, inputs=[query_input, sector_dropdown], outputs=results_output)
     query_input.submit(fn=recommend_investment, inputs=[query_input, sector_dropdown], outputs=results_output)
 
-if __name__ == "__main__":
+if _name_ == "_main_":
     print("Launching SectorSync AI Recommender App...")
+    # share=False is best for Hugging Face spaces
     demo.launch(share=False)