Update app.py

app.py

@@ -119,7 +119,12 @@ def generate_cmx_edl(edl_title, segments, source_name, fps=25):
         edl_lines.append(f"{i:03}  {reel_id}       V     C        {src_in} {src_out} {rec_in} {rec_out}")
         edl_lines.append(f"* FROM CLIP NAME: {source_name}")
         edl_lines.append(f"* {seg.get('note', 'Clip')}\n")
-        rec_start += duration
+        
+        # --- NEW LOGIC: Support Gaps ---
+        # Add the duration of this clip AND any requested gap to the record timeline
+        gap = seg.get('gap', 0.0)
+        rec_start += duration + gap
+        
     return "\n".join(edl_lines)
 
 def call_gemini_for_edl(transcript_data, story_prompt, api_key):
@@ -129,15 +134,19 @@ def call_gemini_for_edl(transcript_data, story_prompt, api_key):
         
     url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-preview-09-2025:generateContent?key={api_key}"
     
+    # --- UPDATED SYSTEM PROMPT ---
     system_prompt = (
         "You are an expert Documentary Senior Editor. Use the provided transcript JSON "
         "(which includes Speaker IDs and word-level timestamps) to create a condensed story. "
-        "Output ONLY a valid JSON array of segments with 'src_start', 'src_end', and 'note'. "
+        "Output ONLY a valid JSON array of segments with 'src_start', 'src_end', 'note', and optionally 'gap'. "
         "CRITICAL RULES:\n"
         "1. IGNORE ALL INTERVIEWER COMMENTS: Do not include any speech or segments where the interviewer is speaking.\n"
         "2. REMOVE FLUFF: Delete 'um', 'ah', repeats, and irrelevant filler.\n"
         "3. NARRATIVE FLOW: Focus on the subject's high-energy responses and narrative hooks.\n"
-        "4. TIMESTAMP INTEGRITY: Use only the exact word-level start and end times from the data."
+        "4. TIMESTAMP INTEGRITY: Use only the exact word-level start and end times from the data.\n"
+        "5. PACING: Group related clips together so they flow seamlessly. HOWEVER, between distinct ideas or sections, "
+        "add a 'gap': 1.0 property (float, in seconds) to the segment preceding the break. "
+        "This will leave blank space on the timeline to signal a topic change."
     )
     
     prompt_text = f"Creative Brief: {story_prompt}\n\nTranscript Data:\n{json.dumps(transcript_data)}"
@@ -219,17 +228,14 @@ if uploaded_file:
         if not ACTIVE_HF_TOKEN or "PASTE_YOUR_HF_TOKEN" in ACTIVE_HF_TOKEN:
             st.error("Please provide a valid Hugging Face Token in the Sidebar/Secrets.")
         else:
-            # Create a dedicated container for the progress UI so it stays at the top
             progress_container = st.container()
             with progress_container:
                 st.info("🤖 **Junior Editor is processing your file...**")
-                
-                # Create the visual elements
                 status_text = st.empty()
                 progress_bar = st.progress(0)
                 
                 try:
-                    # Phase 1: Save File & Extract Audio
+                    # Phase 1: Save File
                     status_text.markdown("**Phase 1/4: Extracting Audio...**")
                     with open("temp_input", "wb") as f:
                         f.write(uploaded_file.getbuffer())
@@ -241,7 +247,6 @@ if uploaded_file:
                     ])
                     progress_bar.progress(25)
                     
-                    # Setup Device
                     device = "cuda" if torch.cuda.is_available() else "cpu"
                     if device == "cpu":
                         st.warning("⚠️ No GPU detected. This will be slow.")
@@ -253,11 +258,8 @@ if uploaded_file:
                     model = whisperx.load_model(model_size, device, compute_type=compute_type)
                     audio = whisperx.load_audio("temp_audio.wav")
                     
-                    # This step takes the most time. It will sit at 25% until complete.
-                    # This is safer than chunking manually which can cause transcription errors.
                     result = model.transcribe(audio, batch_size=16, language=target_language)
                     
-                    # Cleanup VRAM
                     del model
                     gc.collect()
                     torch.cuda.empty_cache()