Update app.py

app.py

@@ -8,19 +8,20 @@ import pytesseract
 import torch
 import os
 import spaces
+import re
 
-# 🔐 Authenticate Hugging Face token
+# 🔐 Hugging Face authentication
 login(token=os.environ.get("token"))
 
-# ✅ Ensure GPU is available
+# ✅ Check GPU availability
 if not torch.cuda.is_available():
     raise RuntimeError("❌ GPU not detected! Please enable GPU in Space settings.")
 print(f"✅ Using GPU: {torch.cuda.get_device_name(0)}")
 
-# 🧠 Model
+# 🧠 Model ID
 model_id = "mistralai/Mistral-7B-Instruct-v0.2"
 
-# 📄 Document extractors
+# 📄 Extractor for PDF (with OCR) and DOCX
 def extract_text_from_pdf(file):
     text = ""
     with pdfplumber.open(file) as pdf:
@@ -51,76 +52,82 @@ def chunk_text(text, max_chars=6000):
         chunks.append(current_chunk)
     return chunks
 
-# 🧾 Q&A Prompt Template
+# 🧾 Prompt optimized for 20 Q&A
 def create_prompt(text_chunk):
     return f"""
 You are an expert in analyzing U.S. government tender documents. Based only on the content provided below, answer the following 20 standard questions in Q&A format. If something is not mentioned, write "Not mentioned in the provided document."
-
 CONTENT:
 {text_chunk}
-
 Now provide answers for:
-
-Q1: What is the general scope of the tender?  
-Q2: Are certifications like SAM, CMMI, ISO, SBA, 8(a), or GSA required?  
-Q3: Is there a Set-aside status (e.g., 8a, SDVOSB)?  
-Q4: Are U.S. citizens or security-cleared staff required?  
-Q5: What is the expected team size or key qualifications?  
-Q6: Are offshore resources allowed?  
-Q7: What is the mode of working (On-site/Remote/Hybrid)?  
-Q8: Is presence in specific regions/states required?  
-Q9: Is the delivery location defined?  
-Q10: Is remote or offshore delivery allowed?  
-Q11: Is a U.S. office presence required?  
-Q12: Are travel/lodging expenses reimbursable?  
-Q13: Are cybersecurity frameworks (FedRAMP, NIST, HIPAA) required?  
-Q14: Are background checks or security clearance needed?  
-Q15: Is past experience required?  
-Q16: How many references are required?  
-Q17: Are only U.S. references accepted?  
-Q18: Is private sector experience allowed?  
-Q19: Do references need to be identified?  
-Q20: Is subcontracting permitted?
-
+1. What is the general scope of the tender?  
+2. Are certifications like SAM, CMMI, ISO, SBA, 8(a), or GSA required?  
+3. Is there a Set-aside status (e.g., 8a, SDVOSB)?  
+4. Are U.S. citizens or security-cleared staff required?  
+5. What is the expected team size or key qualifications?  
+6. Are offshore resources allowed?  
+7. What is the mode of working (On-site/Remote/Hybrid)?  
+8. Is presence in specific regions/states required?  
+9. Is the delivery location defined?  
+10. Is remote or offshore delivery allowed?  
+11. Is a U.S. office presence required?  
+12. Are travel/lodging expenses reimbursable?  
+13. Are cybersecurity frameworks (FedRAMP, NIST, HIPAA) required?  
+14. Are background checks or security clearance needed?  
+15. Is past experience required?  
+16. How many references are required?  
+17. Are only U.S. references accepted?  
+18. Is private sector experience allowed?  
+19. Do references need to be identified?  
+20. Is subcontracting permitted?
+Answer in this format:
+Q1: ...
+A1: ...
+Q2: ...
+A2: ...
 ...
 """
 
-# 🧼 Cleaner
+# ✅ Clean model output to remove repeated prompt content
 def clean_output(raw_output):
-    # Find first valid Q1
-    start_idx = raw_output.find("Q1:")
-    if start_idx == -1:
-        return raw_output.strip()
+    lines = raw_output.splitlines()
+    cleaned = []
+    capture = False
+
+    for line in lines:
+        if line.strip().startswith("Q1:"):
+            capture = True
+        if capture:
+            cleaned.append(line)
 
-    cleaned = raw_output[start_idx:]
+    text = "\n".join(cleaned)
 
-    # Remove everything after second instance of Q1 (to drop repeated prompts)
-    second_q1 = cleaned.find("Q1:", 3)  # skip first one
-    if second_q1 != -1:
-        cleaned = cleaned[:second_q1]
+    # Remove any repeated question block after A20
+    if "Q20:" in text:
+        text = text.split("Q20:")[0] + "Q20: Is subcontracting permitted?"
 
-    # Drop leftover instructions if they show up later
-    cut_phrases = ["You are an expert", "Now provide answers", "CONTENT:", "Answer clearly and in the same format:"]
-    for phrase in cut_phrases:
-        if phrase in cleaned:
-            cleaned = cleaned.split(phrase)[0]
+    # Trim content after A20 if any
+    match = re.search(r"(A20:.*?)\n", text, re.DOTALL)
+    if match:
+        end = match.end()
+        text = text[:end].strip()
 
-    return cleaned.strip()
+    return text.strip()
 
-# 🚀 Main analysis function
+# 🔍 GPU-enabled analyzer
 @spaces.GPU(duration=150)
-def analyze_document(file, cancel_flag):
-    ext = os.path.splitext(file.name)[-1].lower()
+def analyze_document(file, status_text, cancel_flag):
+    filename = file.name
+    ext = os.path.splitext(filename)[-1].lower()
 
     if ext == ".pdf":
         raw_text = extract_text_from_pdf(file)
     elif ext == ".docx":
         raw_text = extract_text_from_docx(file)
     else:
-        return "❌ Unsupported file format. Please upload a PDF or DOCX.", "❌ Invalid format"
+        return "❌ Unsupported file format. Please upload a PDF or DOCX."
 
     if len(raw_text.strip()) == 0:
-        return "❌ No text found in the document.", "❌ Empty document"
+        return "❌ No text found in the document."
 
     chunks = chunk_text(raw_text)
     full_summary = ""
@@ -136,16 +143,17 @@ def analyze_document(file, cancel_flag):
     generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
 
     for i, chunk in enumerate(chunks):
-        if cancel_flag:
-            return "⛔ Analysis cancelled by user.", "⛔ Terminated by user"
+        if cancel_flag.value:
+            return "⛔ Analysis cancelled by user."
 
-        status_msg = f"🔄 Processing chunk {i+1} of {len(chunks)}..."
+        status_text.value = f"🔄 Processing chunk {i+1} of {len(chunks)}..."
         prompt = create_prompt(chunk)
         result = generator(prompt, max_new_tokens=1024, do_sample=False)[0]["generated_text"]
-        cleaned = clean_output(result)
-        full_summary += cleaned + "\n\n---\n\n"
+        answer = clean_output(result)
+        full_summary += answer + "\n\n---\n\n"
 
-    return full_summary.strip(), "✅ Completed"
+    status_text.value = "✅ Completed"
+    return full_summary.strip()
 
 # 🌐 Gradio Interface
 with gr.Blocks(title="Smart Tender Analyzer - US Edition") as demo:
@@ -157,17 +165,17 @@ with gr.Blocks(title="Smart Tender Analyzer - US Edition") as demo:
             with gr.Row():
                 analyze_button = gr.Button("🔍 Analyze", variant="primary")
                 terminate_button = gr.Button("❌ Terminate", variant="stop")
-            status_box = gr.Textbox(label="📊 Status", value="⏳ Waiting for input...", interactive=False)
+            status_text = gr.Textbox(label="📊 Status", value="⏳ Waiting for input...", interactive=False)
 
         with gr.Column(scale=2):
             output_box = gr.Textbox(label="🧠 Extracted Tender Intelligence", lines=30, interactive=False)
 
-    cancel_flag = gr.State(False)
+    cancel_flag = gr.State(value=False)
 
     analyze_button.click(
         fn=analyze_document,
-        inputs=[file_input, cancel_flag],
-        outputs=[output_box, status_box]
+        inputs=[file_input, status_text, cancel_flag],
+        outputs=output_box
     )
 
     terminate_button.click(