Update app.py

app.py

@@ -153,7 +153,7 @@ def extract_youtube_audio(url, max_duration_sec=300, cookiefile=None):
     
     return wav_path, title
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=180)
 def transcribe_audio_with_asr(audio_path):
     """Transcribe audio file using Cohere ASR model via transformers."""
     global asr_model, asr_processor
@@ -168,6 +168,14 @@ def transcribe_audio_with_asr(audio_path):
     texts = asr_processor.decode(outputs, skip_special_tokens=True)
     # text = texts[0] if isinstance(texts, list) else texts
     # join texts
+    # Filter the lines in texts which are english only and no korean
+    if isinstance(texts, list):
+        # Filter out lines that are purely English/symbols (no Korean characters)
+        # Korean Unicode range: AC00-D7A3 (Syllables), 1100-11FF (Jamo), 3130-318F (Compatibility Jamo)
+        korean_re = re_module.compile(r'[가-힣ㄱ-ㅎㅏ-ㅣ]')
+        texts = [line for line in texts if korean_re.search(line)]
+        
+
     text = "\n".join(texts) if isinstance(texts, list) else texts
     
     
@@ -241,7 +249,7 @@ def get_base64_image(image):
     img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
     return f"data:image/jpeg;base64,{img_str}"
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=180)
 def extract_vocabulary(pdf_text, images, translit_lang, translit_format, target_lang, max_text_char=1500, repetition_penalty_val=1.1, partial_assistant_text=None, auto_force_chars=1000):
     """Use Transformers to extract vocabulary from text and images."""
     global model, processor
@@ -249,32 +257,54 @@ def extract_vocabulary(pdf_text, images, translit_lang, translit_format, target_
     os.makedirs("log", exist_ok=True)
     
     if len(pdf_text.strip()) == 0:
-        pdf_text = '''"No Text available, see provided images only."'''
-
+        pdf_text = '''"No Text available, see provided Images only."'''
+    
+    no_img = ""
+    if len(images) == 0:
+        no_img = '''\n"No Images available, see provided Text only."'''
 
     non_english = ""
     if translit_lang.upper() != "ENGLISH":
         non_english = f" CRITICAL: You MUST use the native alphabet/script of {translit_lang.upper()}, do NOT use English letters unless requested."
     
-    prompt_text = f"""Extract at least 10 key Korean words or phrases from the following text and images.
+    prompt_text = f"""Text:
+
+<scrpated-content>
+{pdf_text[:int(max_text_char)]}
+</scrpated-content>{no_img}
+
+Extract at least 10 key Korean words or phrases from the following text and images.
 Focus on meaningful vocabulary that is highly helpful for a new language learner (e.g., common nouns, verbs, adjectives, or useful expressions).
 CRITICAL: Do NOT extract website template words, navigation menus, boilerplate text, UI elements, or titles like 'Home page', 'News', 'Menu'.
 
 Return ONLY a valid JSON list of dictionaries, where each dictionary has four keys:
 - 'korean' (the Korean text)
 - 'transliteration' (the pronunciation transliterated into {translit_lang.upper()} script/characters, formatted as {translit_format}.{non_english})
-- 'translation' (the translation into {target_lang.upper()})
+- 'translation' (the brief translation into {target_lang.upper()})
 - 'explanation' (a brief grammar or context note in {target_lang.upper()}).
 
-Just output raw JSON with ```json and ``` markers, as the user will load in python.
+Just output raw JSON with ```json and ``` markers, as the user will load in python. Example:
+
+```json
+[
+  {{
+    "korean": "날씨",
+    "transliteration": "nal-ssi",
+    "translation": "weather",
+    "explanation": "Common noun used to describe weather conditions."
+  }},
+  {{
+    "korean": "맛있다",
+    "transliteration": "ma-sit-da",
+    "translation": "to be delicious",
+    "explanation": "Descriptive verb. Polite form: 맛있어요. Used to compliment food."
+  }}
+]
+```
 
 CRITICAL: Do NOT overthink. Do NOT deliberate over conditions, edge cases, or reasoning. Keep your thinking extremely brief (a few words at most). Output the JSON array IMMEDIATELY without lengthy analysis.
 
-Text:
 
-<scrpated-content>
-{pdf_text[:int(max_text_char)]}
-</scrpated-content>
 """
     
     # DEBUG: Log prompt text
@@ -375,7 +405,7 @@ Text:
             # Check if user clicked "Stop thinking" OR auto-force threshold reached
             if (global_stop_thinking[0] or should_auto_force) and not force_triggered:
                 force_triggered = True
-                reason = "auto-force (>300 chars)" if should_auto_force else "user clicked stop"
+                reason = f"auto-force (>{AUTO_FORCE_CHARS} chars)" if should_auto_force else "user clicked stop"
                 print(f"[STOP-THINK] Force triggered ({reason})! Killing current generation...")
                 
                 # 1. Kill the current generation thread
@@ -482,8 +512,13 @@ Text:
             # Fallback: find last [ ... ] or { ... } block
             json_matches = list(re.finditer(r'(\[[\s\S]*\]|\{[\s\S]*\})', output_text))
             clean_text = json_matches[-1].group(1).strip() if json_matches else output_text.strip()
-            
-        data = json.loads(clean_text)
+        try:
+            data = json.loads(clean_text)
+        except:
+            import jiter
+            # Get bytes from string
+            data = jiter.from_json(clean_text.encode("utf-8"), partial_mode=True)
+        
         if not isinstance(data, list):
             data = [data]
         yield output_text, data
@@ -568,7 +603,12 @@ Korean words:
             json_matches = list(re.finditer(r'(\[[\s\S]*\]|\{[\s\S]*\})', output_text))
             clean_text = json_matches[-1].group(1).strip() if json_matches else output_text.strip()
             
-        data = json.loads(clean_text)
+        try:
+            data = json.loads(clean_text)
+        except:
+            import jiter
+            data = jiter.from_json(clean_text.encode("utf-8"), partial_mode=True)
+            
         if not isinstance(data, list):
             data = [data]
         return data
@@ -590,7 +630,7 @@ def hash_file(filepath):
     with open(filepath, 'rb') as f:
         return hashlib.md5(f.read(1024*1024)).hexdigest()
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=180)
 def process_pdf(pdf_file, url_input, audio_file_input, yt_url_input, yt_cookies_file, translit_lang, translit_format, target_lang, max_text_char, repetition_penalty_val, auto_force_chars_val, last_source_hash, last_korean_words, active_tab, progress=gr.Progress()):
     global tts, voice_style
     
@@ -705,7 +745,11 @@ def process_pdf(pdf_file, url_input, audio_file_input, yt_url_input, yt_cookies_
                 clean_text = json_matches[-1].group(1).strip() if json_matches else ""
             
             if clean_text:
-                data = json.loads(clean_text)
+                try:
+                    data = json.loads(clean_text)
+                except:
+                    import jiter
+                    data = jiter.from_json(clean_text.encode("utf-8"), partial_mode=True)
                 if not isinstance(data, list):
                     data = [data]
                 if data and isinstance(data[0], dict) and 'korean' in data[0]:
@@ -1083,7 +1127,7 @@ def get_example_audio():
             print(f"Failed to download example audio: {e}")
     return file_path if os.path.exists(file_path) else None
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=180)
 def process_pdf_force(partial_text, pdf_file, url_input, translit_lang, translit_format, target_lang, max_text_char, repetition_penalty_val, last_source_state, last_korean_words_state):
     """Force JSON generation using the current partial stream_box text."""
     is_url = bool(url_input and url_input.strip())
@@ -1309,7 +1353,7 @@ def create_demo():
                 gr.Markdown("### ⚙️ Customization Settings")
                 max_text_char_input = gr.Slider(minimum=1000, maximum=30000, step=1000, value=1500, label="Max Input Text Length (Characters)")
                 repetition_penalty_input = gr.Slider(minimum=0.1, maximum=2.0, step=0.1, value=1.2, label="Repetition Penalty")
-                auto_force_chars_input = gr.Slider(minimum=100, maximum=5000, step=100, value=1000, label="Auto-force JSON after (chars of thinking)")
+                auto_force_chars_input = gr.Slider(minimum=1_000, maximum=10_000, step=100, value=4_000, label="Auto-force JSON after (chars of thinking)")
                 
                 with gr.Accordion("🔧 Advanced", open=False):
                     translit_lang = gr.Dropdown(