Spaces:
				
			
			
	
			
			
		Runtime error
		
	
	
	
			
			
	
	
	
	
		
		
		Runtime error
		
	Update main.py
Browse files
    	
        main.py
    CHANGED
    
    | 
         @@ -1,4 +1,4 @@ 
     | 
|
| 1 | 
         
            -
            # main.py
         
     | 
| 2 | 
         
             
            import os
         
     | 
| 3 | 
         
             
            os.environ['HF_HOME'] = '/tmp'
         
     | 
| 4 | 
         | 
| 
         @@ -22,27 +22,26 @@ logging.basicConfig(level=logging.INFO) 
     | 
|
| 22 | 
         
             
            logger = logging.getLogger("summarizer")
         
     | 
| 23 | 
         | 
| 24 | 
         
             
            # -------------------------
         
     | 
| 25 | 
         
            -
            # Device selection (CPU 
     | 
| 26 | 
         
             
            # -------------------------
         
     | 
| 27 | 
         
            -
            # The user asked for CPU — force device -1. If you later enable GPU, set DEVICE accordingly.
         
     | 
| 28 | 
         
             
            USE_GPU = False
         
     | 
| 29 | 
         
             
            DEVICE = -1
         
     | 
| 30 | 
         
            -
            logger.info("Startup: forcing CPU usage for  
     | 
| 31 | 
         | 
| 32 | 
         
             
            # -------------------------
         
     | 
| 33 | 
         
             
            # Model names and caches
         
     | 
| 34 | 
         
             
            # -------------------------
         
     | 
| 35 | 
         
             
            PEGASUS_MODEL = "google/pegasus-large"
         
     | 
| 36 | 
         
             
            LED_MODEL = "allenai/led-large-16384"
         
     | 
| 37 | 
         
            -
             
     | 
| 38 | 
         
            -
            PARAM_MODEL = "google/flan-t5-small" 
     | 
| 39 | 
         | 
| 40 | 
         
             
            _SUMMARIZER_CACHE: Dict[str, Any] = {}
         
     | 
| 41 | 
         
             
            _PARAM_GENERATOR = None
         
     | 
| 42 | 
         
             
            _PREFERRED_SUMMARIZER_KEY: Optional[str] = None
         
     | 
| 43 | 
         | 
| 44 | 
         
             
            # -------------------------
         
     | 
| 45 | 
         
            -
            # Utilities
         
     | 
| 46 | 
         
             
            # -------------------------
         
     | 
| 47 | 
         
             
            _STOPWORDS = {
         
     | 
| 48 | 
         
             
                "the","and","is","in","to","of","a","that","it","on","for","as","are","with","was","be","by","this","an","or","from","at","which","we","has","have"
         
     | 
| 
         @@ -68,7 +67,6 @@ def extractive_prefilter(text: str, top_k: int = 6) -> str: 
     | 
|
| 68 | 
         
             
                return " ".join(chosen)
         
     | 
| 69 | 
         | 
| 70 | 
         
             
            def chunk_text_by_chars(text: str, max_chars: int = 800, overlap: int = 120) -> List[str]:
         
     | 
| 71 | 
         
            -
                # small chunks to keep CPU generation per-call bounded
         
     | 
| 72 | 
         
             
                n = len(text)
         
     | 
| 73 | 
         
             
                if n <= max_chars:
         
     | 
| 74 | 
         
             
                    return [text]
         
     | 
| 
         @@ -89,97 +87,98 @@ def chunk_text_by_chars(text: str, max_chars: int = 800, overlap: int = 120) -> 
     | 
|
| 89 | 
         
             
                        start = end
         
     | 
| 90 | 
         
             
                return [p for p in parts if p]
         
     | 
| 91 | 
         | 
| 92 | 
         
            -
            def _first_int_from_text(s: str, fallback: Optional[int] = None) -> Optional[int]:
         
     | 
| 93 | 
         
            -
                m = re.search(r"\d{1,4}", s)
         
     | 
| 94 | 
         
            -
                return int(m.group()) if m else fallback
         
     | 
| 95 | 
         
            -
             
     | 
| 96 | 
         
             
            # -------------------------
         
     | 
| 97 | 
         
            -
            #  
     | 
| 98 | 
         
             
            # -------------------------
         
     | 
| 99 | 
         
            -
            def  
     | 
| 100 | 
         
            -
                 
     | 
| 101 | 
         
            -
                 
     | 
| 102 | 
         
            -
             
     | 
| 103 | 
         
            -
             
     | 
| 104 | 
         
            -
             
     | 
| 105 | 
         
            -
             
     | 
| 106 | 
         
            -
                    _PARAM_GENERATOR = pipeline("text2text-generation", model=p_mod, tokenizer=p_tok, device=DEVICE)
         
     | 
| 107 | 
         
            -
                    logger.info("Param-generator loaded as text2text-generation.")
         
     | 
| 108 | 
         
            -
                except Exception as e:
         
     | 
| 109 | 
         
            -
                    logger.exception("Param-generator failed to load as text2text: %s", e)
         
     | 
| 110 | 
         
            -
                    _PARAM_GENERATOR = None
         
     | 
| 111 | 
         
            -
             
     | 
| 112 | 
         
            -
            def preload_models_at_startup():
         
     | 
| 113 | 
         
            -
                global _PREFERRED_SUMMARIZER_KEY
         
     | 
| 114 | 
         
            -
                # Preload Pegasus and LED (best-effort). If fail, load fallback.
         
     | 
| 115 | 
         
            -
                logger.info("Preloading summarizer models (Pegasus, LED, fallback)...")
         
     | 
| 116 | 
         
            -
                p = safe_load_pipeline(PEGASUS_MODEL)
         
     | 
| 117 | 
         
            -
                if p:
         
     | 
| 118 | 
         
            -
                    _SUMMARIZER_CACHE["pegasus"] = p
         
     | 
| 119 | 
         
            -
                    _PREFERRED_SUMMARIZER_KEY = "pegasus"
         
     | 
| 120 | 
         
            -
                else:
         
     | 
| 121 | 
         
            -
                    logger.warning("Pegasus failed to load on CPU; will still attempt LED and fallback.")
         
     | 
| 122 | 
         
            -
             
     | 
| 123 | 
         
            -
                l = safe_load_pipeline(LED_MODEL)
         
     | 
| 124 | 
         
            -
                if l:
         
     | 
| 125 | 
         
            -
                    _SUMMARIZER_CACHE["led"] = l
         
     | 
| 126 | 
         
            -
             
     | 
| 127 | 
         
            -
                # Always ensure fallback available
         
     | 
| 128 | 
         
            -
                f = safe_load_pipeline(FALLBACK_MODEL)
         
     | 
| 129 | 
         
            -
                if f:
         
     | 
| 130 | 
         
            -
                    _SUMMARIZER_CACHE["distilbart"] = f
         
     | 
| 131 | 
         
            -
                    if not _PREFERRED_SUMMARIZER_KEY:
         
     | 
| 132 | 
         
            -
                        _PREFERRED_SUMMARIZER_KEY = "distilbart"
         
     | 
| 133 | 
         
            -
                else:
         
     | 
| 134 | 
         
            -
                    logger.critical("Fallback model failed to load. The app will not be able to summarize.")
         
     | 
| 135 | 
         
            -
             
     | 
| 136 | 
         
            -
                # Load param generator if possible (small model)
         
     | 
| 137 | 
         
             
                try:
         
     | 
| 138 | 
         
            -
                     
     | 
| 139 | 
         
            -
                     
     | 
| 140 | 
         
            -
                     
     | 
| 141 | 
         
            -
             
     | 
| 142 | 
         
            -
             
     | 
| 143 | 
         
            -
                     
     | 
| 144 | 
         
            -
             
     | 
| 145 | 
         
            -
             
     | 
| 146 | 
         
            -
                     
     | 
| 147 | 
         
            -
             
     | 
| 148 | 
         
            -
             
     | 
| 149 | 
         
            -
             
     | 
| 150 | 
         
            -
             
     | 
| 151 | 
         
            -
             
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 152 | 
         | 
| 153 | 
         
             
            # -------------------------
         
     | 
| 154 | 
         
            -
            #  
     | 
| 155 | 
         
             
            # -------------------------
         
     | 
| 156 | 
         
            -
            def  
     | 
| 157 | 
         
             
                """
         
     | 
| 158 | 
         
            -
                 
     | 
| 159 | 
         
            -
                 
     | 
| 160 | 
         
             
                """
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 161 | 
         
             
                model_name = getattr(pipe.model.config, "name_or_path", "") or ""
         
     | 
| 162 | 
         
            -
                 
     | 
| 163 | 
         
            -
                is_led = "led" in model_name_lower or "longformer" in model_name_lower
         
     | 
| 164 | 
         | 
| 165 | 
         
            -
                #  
     | 
| 166 | 
         
             
                fast_cfg = {
         
     | 
| 167 | 
         
             
                    "max_new_tokens": 64 if short_target else (120 if not is_led else 240),
         
     | 
| 168 | 
         
             
                    "do_sample": True,
         
     | 
| 169 | 
         
            -
                    "top_p": 0. 
     | 
| 170 | 
         
             
                    "temperature": 0.85,
         
     | 
| 171 | 
         
             
                    "num_beams": 1,
         
     | 
| 172 | 
         
             
                    "early_stopping": True,
         
     | 
| 173 | 
         
             
                    "no_repeat_ngram_size": 3,
         
     | 
| 174 | 
         
             
                }
         
     | 
| 175 | 
         
            -
             
     | 
| 176 | 
         
             
                try:
         
     | 
| 177 | 
         
            -
                     
     | 
| 178 | 
         
            -
                    return out
         
     | 
| 179 | 
         
             
                except Exception as e:
         
     | 
| 180 | 
         
            -
                    logger.warning("Fast pass failed: %s 
     | 
| 181 | 
         | 
| 182 | 
         
            -
                # QUALITY beam pass (more expensive)
         
     | 
| 183 | 
         
             
                quality_cfg = {
         
     | 
| 184 | 
         
             
                    "max_new_tokens": 140 if not is_led else 320,
         
     | 
| 185 | 
         
             
                    "do_sample": False,
         
     | 
| 
         @@ -187,121 +186,99 @@ def summarize_with_model(pipe, text_prompt: str, short_target: bool = False) -> 
     | 
|
| 187 | 
         
             
                    "early_stopping": True,
         
     | 
| 188 | 
         
             
                    "no_repeat_ngram_size": 3,
         
     | 
| 189 | 
         
             
                }
         
     | 
| 190 | 
         
            -
             
     | 
| 191 | 
         
             
                try:
         
     | 
| 192 | 
         
            -
                     
     | 
| 193 | 
         
            -
             
     | 
| 194 | 
         
            -
             
     | 
| 195 | 
         
            -
                    logger.exception("Quality pass failed: %s", e2)
         
     | 
| 196 | 
         | 
| 197 | 
         
            -
                # extractive 
     | 
| 198 | 
         
             
                try:
         
     | 
| 199 | 
         
             
                    return extractive_prefilter(text_prompt, top_k=3)
         
     | 
| 200 | 
         
             
                except Exception:
         
     | 
| 201 | 
         
            -
                    return "Summarization failed;  
     | 
| 202 | 
         | 
| 203 | 
         
             
            # -------------------------
         
     | 
| 204 | 
         
            -
            # Param generator (AI decision)  
     | 
| 205 | 
         
             
            # -------------------------
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 206 | 
         
             
            def generate_summarization_config(text: str) -> Dict[str, Any]:
         
     | 
| 207 | 
         
            -
                """
         
     | 
| 208 | 
         
            -
                Uses the text2text param-generator to output a JSON config.
         
     | 
| 209 | 
         
            -
                If the generator fails or returns something noisy (e.g., echoes the input),
         
     | 
| 210 | 
         
            -
                fall back to a safe heuristic.
         
     | 
| 211 | 
         
            -
                """
         
     | 
| 212 | 
         
             
                defaults = {"short": (12, 50), "medium": (50, 130), "long": (130, 300)}
         
     | 
| 213 | 
         
            -
                 
     | 
| 214 | 
         
            -
                 
     | 
| 215 | 
         
             
                    words = len(text.split())
         
     | 
| 216 | 
         
             
                    length = "short" if words < 150 else ("medium" if words < 800 else "long")
         
     | 
| 217 | 
         
             
                    mn, mx = defaults[length]
         
     | 
| 218 | 
         
             
                    return {"length": length, "min_length": mn, "max_length": mx, "tone": "neutral"}
         
     | 
| 219 | 
         | 
| 220 | 
         
            -
                pg = _PARAM_GENERATOR
         
     | 
| 221 | 
         
            -
                if pg is None:
         
     | 
| 222 | 
         
            -
                    logger.info("Param-generator not available; using fallback heuristic.")
         
     | 
| 223 | 
         
            -
                    return fallback()
         
     | 
| 224 | 
         
            -
             
     | 
| 225 | 
         
             
                prompt = (
         
     | 
| 226 | 
         
            -
                    "Recommend summarization settings for this text. Answer ONLY with JSON  
     | 
| 227 | 
         
             
                    '{"length":"short|medium|long","tone":"neutral|formal|casual|bullet","min_words":MIN,"max_words":MAX}\n\n'
         
     | 
| 228 | 
         
             
                    "Text:\n'''"
         
     | 
| 229 | 
         
             
                    + text[:3000] + "'''"
         
     | 
| 230 | 
         
             
                )
         
     | 
| 231 | 
         
            -
             
     | 
| 232 | 
         
             
                try:
         
     | 
| 233 | 
         
            -
                    out_item = pg(prompt, max_new_tokens=64, do_sample=False, num_beams=1 
     | 
| 234 | 
         
            -
                     
     | 
| 235 | 
         
            -
                    out = out_item.get("generated_text") or out_item.get("summary_text") or out_item.get("text") or ""
         
     | 
| 236 | 
         
             
                    out = (out or "").strip()
         
     | 
| 237 | 
         
            -
             
     | 
| 238 | 
         
            -
                    # COMMON FAILURE MODE: the model just echoes the input — reject that
         
     | 
| 239 | 
         
            -
                    # If output contains a long substring of the input, treat as invalid.
         
     | 
| 240 | 
         
             
                    if not out:
         
     | 
| 241 | 
         
             
                        raise ValueError("Empty param-generator output")
         
     | 
| 242 | 
         
            -
                    #  
     | 
| 243 | 
         
             
                    input_words = set(w.lower() for w in re.findall(r"\w+", text)[:200])
         
     | 
| 244 | 
         
             
                    out_words = set(w.lower() for w in re.findall(r"\w+", out)[:200])
         
     | 
| 245 | 
         
            -
                    if len(input_words)  
     | 
| 246 | 
         
            -
                        logger.warning("Param-generator appears to echo input;  
     | 
| 247 | 
         
            -
                         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 248 | 
         | 
| 249 | 
         
            -
                    # Find JSON object in output
         
     | 
| 250 | 
         
             
                    jmatch = re.search(r"\{.*\}", out, re.DOTALL)
         
     | 
| 251 | 
         
             
                    if jmatch:
         
     | 
| 252 | 
         
             
                        raw = jmatch.group().replace("'", '"')
         
     | 
| 253 | 
         
             
                        cfg = json.loads(raw)
         
     | 
| 254 | 
         
             
                    else:
         
     | 
| 255 | 
         
            -
                        # attempt to parse line with key:value pairs (tolerant)
         
     | 
| 256 | 
         
             
                        cfg = None
         
     | 
| 257 | 
         | 
| 258 | 
         
             
                    if not cfg or not isinstance(cfg, dict):
         
     | 
| 259 | 
         
            -
                         
     | 
| 260 | 
         
            -
             
     | 
| 261 | 
         
            -
             
     | 
| 262 | 
         
            -
                     
     | 
| 263 | 
         
            -
                     
     | 
| 264 | 
         
            -
                    min_w = cfg.get("min_words") or cfg.get("min_length") or cfg.get("min")
         
     | 
| 265 | 
         
            -
                    max_w = cfg.get("max_words") or cfg.get("max_length") or cfg.get("max")
         
     | 
| 266 | 
         
            -
             
     | 
| 267 | 
         
            -
                    if length not in ("short","medium","long"):
         
     | 
| 268 | 
         
            -
                        words = len(text.split())
         
     | 
| 269 | 
         
            -
                        length = "short" if words < 150 else ("medium" if words < 800 else "long")
         
     | 
| 270 | 
         
            -
                    if tone not in ("neutral","formal","casual","bullet"):
         
     | 
| 271 | 
         
            -
                        tone = "neutral"
         
     | 
| 272 | 
         
            -
             
     | 
| 273 | 
         
            -
                    defaults_min, defaults_max = defaults.get(length, (50,130))
         
     | 
| 274 | 
         
            -
                    try:
         
     | 
| 275 | 
         
            -
                        mn = int(min_w) if min_w is not None else defaults_min
         
     | 
| 276 | 
         
            -
                        mx = int(max_w) if max_w is not None else defaults_max
         
     | 
| 277 | 
         
            -
                    except Exception:
         
     | 
| 278 | 
         
            -
                        mn, mx = defaults_min, defaults_max
         
     | 
| 279 | 
         
            -
             
     | 
| 280 | 
         
             
                    mn = max(5, min(mn, 2000))
         
     | 
| 281 | 
         
             
                    mx = max(mn + 5, min(mx, 4000))
         
     | 
| 282 | 
         
            -
                    logger.info("Param-generator suggested length=%s tone=%s min=%s max=%s", length, tone, mn, mx)
         
     | 
| 283 | 
         
             
                    return {"length": length, "min_length": mn, "max_length": mx, "tone": tone}
         
     | 
| 284 | 
         
            -
             
     | 
| 285 | 
         
             
                except Exception as e:
         
     | 
| 286 | 
         
            -
                    logger.exception("Param-generator failed 
     | 
| 287 | 
         
            -
                     
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 288 | 
         | 
| 289 | 
         
             
            # -------------------------
         
     | 
| 290 | 
         
            -
            #  
     | 
| 291 | 
         
             
            # -------------------------
         
     | 
| 292 | 
         
            -
             
     | 
| 293 | 
         
            -
             
     | 
| 294 | 
         
            -
             
     | 
| 295 | 
         
            -
            REFINE_TIMEOUT_SECONDS = 60  # final refinement timeout
         
     | 
| 296 | 
         
            -
            MAX_TOTAL_SECONDS = 180      # overall safety cap for a request
         
     | 
| 297 | 
         
            -
             
     | 
| 298 | 
         
            -
            executor = ThreadPoolExecutor(max_workers=MAX_WORKERS)
         
     | 
| 299 | 
         | 
| 300 | 
         
             
            def summarize_chunks_parallel(pipe, chunks: List[str], chunk_target: int) -> List[str]:
         
     | 
| 301 | 
         
            -
                """
         
     | 
| 302 | 
         
            -
                Submit chunk summarization tasks to threadpool; apply per-chunk timeout.
         
     | 
| 303 | 
         
            -
                If a chunk times out or fails, use extractive_prefilter fallback.
         
     | 
| 304 | 
         
            -
                """
         
     | 
| 305 | 
         
             
                futures = {}
         
     | 
| 306 | 
         
             
                results = [None] * len(chunks)
         
     | 
| 307 | 
         
             
                for idx, chunk in enumerate(chunks):
         
     | 
| 
         @@ -310,26 +287,24 @@ def summarize_chunks_parallel(pipe, chunks: List[str], chunk_target: int) -> Lis 
     | 
|
| 310 | 
         
             
                    futures[fut] = idx
         
     | 
| 311 | 
         | 
| 312 | 
         
             
                start = time.time()
         
     | 
| 313 | 
         
            -
                for fut in as_completed(futures 
     | 
| 314 | 
         
             
                    idx = futures[fut]
         
     | 
| 315 | 
         
             
                    try:
         
     | 
| 316 | 
         
             
                        remaining = max(0.1, CHUNK_TIMEOUT_SECONDS - (time.time() - start))
         
     | 
| 317 | 
         
            -
                        # bound waiting for each future to CHUNK_TIMEOUT_SECONDS
         
     | 
| 318 | 
         
             
                        results[idx] = fut.result(timeout=remaining)
         
     | 
| 319 | 
         
             
                    except TimeoutError:
         
     | 
| 320 | 
         
             
                        logger.warning("Chunk %d timed out; using extractive fallback.", idx)
         
     | 
| 321 | 
         
             
                        results[idx] = extractive_prefilter(chunks[idx], top_k=3)
         
     | 
| 322 | 
         
             
                    except Exception as e:
         
     | 
| 323 | 
         
            -
                        logger.exception("Chunk %d failed: %s;  
     | 
| 324 | 
         
             
                        results[idx] = extractive_prefilter(chunks[idx], top_k=3)
         
     | 
| 325 | 
         
            -
                # ensure no None
         
     | 
| 326 | 
         
             
                for i, r in enumerate(results):
         
     | 
| 327 | 
         
             
                    if not r:
         
     | 
| 328 | 
         
             
                        results[i] = extractive_prefilter(chunks[i], top_k=3)
         
     | 
| 329 | 
         
             
                return results
         
     | 
| 330 | 
         | 
| 331 | 
         
             
            # -------------------------
         
     | 
| 332 | 
         
            -
            # Prompt  
     | 
| 333 | 
         
             
            # -------------------------
         
     | 
| 334 | 
         
             
            def apply_tone_instruction(text: str, tone: str, target_sentences: Optional[int] = None) -> str:
         
     | 
| 335 | 
         
             
                tone = (tone or "neutral").lower()
         
     | 
| 
         @@ -337,13 +312,13 @@ def apply_tone_instruction(text: str, tone: str, target_sentences: Optional[int] 
     | 
|
| 337 | 
         
             
                    instr = "Produce concise bullet points. Each bullet <= 20 words. No extra commentary."
         
     | 
| 338 | 
         
             
                elif tone == "short":
         
     | 
| 339 | 
         
             
                    ts = target_sentences or 1
         
     | 
| 340 | 
         
            -
                    instr = f"Summarize in {ts} sentence{'s' if ts>1 else ''}. Be abstractive 
     | 
| 341 | 
         
             
                elif tone == "formal":
         
     | 
| 342 | 
         
            -
                    instr = "Summarize in a formal professional tone  
     | 
| 343 | 
         
             
                elif tone == "casual":
         
     | 
| 344 | 
         
            -
                    instr = "Summarize in a casual tone  
     | 
| 345 | 
         
             
                elif tone == "long":
         
     | 
| 346 | 
         
            -
                    instr = "Provide  
     | 
| 347 | 
         
             
                else:
         
     | 
| 348 | 
         
             
                    instr = "Summarize in 2-3 clear sentences."
         
     | 
| 349 | 
         
             
                instr += " Do not repeat information. Prefer rephrasing."
         
     | 
| 
         @@ -354,30 +329,51 @@ def refine_combined(pipe, summaries_list: List[str], tone: str, final_target_sen 
     | 
|
| 354 | 
         
             
                if len(combined.split()) > 1200:
         
     | 
| 355 | 
         
             
                    combined = extractive_prefilter(combined, top_k=20)
         
     | 
| 356 | 
         
             
                prompt = apply_tone_instruction(combined, tone, target_sentences=final_target_sentences)
         
     | 
| 357 | 
         
            -
                # run refine in executor to apply timeout
         
     | 
| 358 | 
         
             
                fut = executor.submit(summarize_with_model, pipe, prompt, short_target=False)
         
     | 
| 359 | 
         
             
                try:
         
     | 
| 360 | 
         
             
                    return fut.result(timeout=REFINE_TIMEOUT_SECONDS)
         
     | 
| 361 | 
         
             
                except TimeoutError:
         
     | 
| 362 | 
         
            -
                    logger.warning("Refine  
     | 
| 363 | 
         
             
                    return " ".join(summaries_list[:6])
         
     | 
| 364 | 
         
             
                except Exception as e:
         
     | 
| 365 | 
         
            -
                    logger.exception("Refine  
     | 
| 366 | 
         
             
                    return " ".join(summaries_list[:6])
         
     | 
| 367 | 
         | 
| 368 | 
         
             
            # -------------------------
         
     | 
| 369 | 
         
            -
            #  
     | 
| 370 | 
         
             
            # -------------------------
         
     | 
| 371 | 
         
             
            @app.route("/", methods=["GET"])
         
     | 
| 372 | 
         
             
            def home():
         
     | 
| 373 | 
         
             
                try:
         
     | 
| 374 | 
         
             
                    return render_template("index.html")
         
     | 
| 375 | 
         
             
                except Exception:
         
     | 
| 376 | 
         
            -
                    return "Summarizer ( 
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 377 | 
         | 
| 378 | 
         
             
            @app.route("/summarize", methods=["POST"])
         
     | 
| 379 | 
         
             
            def summarize_route():
         
     | 
| 380 | 
         
            -
                 
     | 
| 381 | 
         
             
                data = request.get_json(force=True) or {}
         
     | 
| 382 | 
         
             
                text = (data.get("text") or "").strip()[:90000]
         
     | 
| 383 | 
         
             
                user_model_pref = (data.get("model") or "auto").lower()
         
     | 
| 
         @@ -387,16 +383,16 @@ def summarize_route(): 
     | 
|
| 387 | 
         
             
                if not text or len(text.split()) < 5:
         
     | 
| 388 | 
         
             
                    return jsonify({"error": "Input too short."}), 400
         
     | 
| 389 | 
         | 
| 390 | 
         
            -
                #  
     | 
| 391 | 
         
            -
                if requested_length in ("auto", 
     | 
| 392 | 
         
             
                    cfg = generate_summarization_config(text)
         
     | 
| 393 | 
         
            -
                    length_choice = cfg.get("length", 
     | 
| 394 | 
         
            -
                    tone_choice = cfg.get("tone", 
     | 
| 395 | 
         
             
                else:
         
     | 
| 396 | 
         
             
                    length_choice = requested_length if requested_length in ("short","medium","long") else "medium"
         
     | 
| 397 | 
         
             
                    tone_choice = requested_tone if requested_tone in ("neutral","formal","casual","bullet") else "neutral"
         
     | 
| 398 | 
         | 
| 399 | 
         
            -
                #  
     | 
| 400 | 
         
             
                words_len = len(text.split())
         
     | 
| 401 | 
         
             
                prefer_led = False
         
     | 
| 402 | 
         
             
                if user_model_pref == "led":
         
     | 
| 
         @@ -406,84 +402,65 @@ def summarize_route(): 
     | 
|
| 406 | 
         
             
                else:
         
     | 
| 407 | 
         
             
                    if length_choice == "long" or words_len > 3000:
         
     | 
| 408 | 
         
             
                        prefer_led = True
         
     | 
| 409 | 
         
            -
                    else:
         
     | 
| 410 | 
         
            -
                        prefer_led = False
         
     | 
| 411 | 
         | 
| 412 | 
         
            -
                # Choose actual pipe
         
     | 
| 413 | 
         
             
                model_key = "led" if prefer_led else (_PREFERRED_SUMMARIZER_KEY or "distilbart")
         
     | 
| 414 | 
         
            -
                 
     | 
| 415 | 
         
            -
                     
     | 
| 416 | 
         
            -
             
     | 
| 417 | 
         
            -
             
     | 
| 418 | 
         
            -
             
     | 
| 419 | 
         
            -
                            _SUMMARIZER_CACHE[model_key] = pipe
         
     | 
| 420 | 
         
            -
                    except Exception as e:
         
     | 
| 421 | 
         
            -
                        logger.exception("On-demand load failed: %s", e)
         
     | 
| 422 | 
         
            -
             
     | 
| 423 | 
         
            -
                # Ensure we have at least a fallback
         
     | 
| 424 | 
         
            -
                if model_key not in _SUMMARIZER_CACHE:
         
     | 
| 425 | 
         
             
                    model_key = "distilbart"
         
     | 
| 426 | 
         
            -
                summarizer_pipe = _SUMMARIZER_CACHE[model_key]
         
     | 
| 427 | 
         | 
| 428 | 
         
            -
                #  
     | 
| 429 | 
         
             
                if model_key != "led" and words_len > 2500:
         
     | 
| 430 | 
         
             
                    text_for_chunks = extractive_prefilter(text, top_k=40)
         
     | 
| 431 | 
         
             
                else:
         
     | 
| 432 | 
         
             
                    text_for_chunks = text
         
     | 
| 433 | 
         | 
| 434 | 
         
            -
                #  
     | 
| 435 | 
         
             
                if model_key == "led":
         
     | 
| 436 | 
         
             
                    chunk_max = 6000
         
     | 
| 437 | 
         
             
                    overlap = 400
         
     | 
| 438 | 
         
             
                else:
         
     | 
| 439 | 
         
            -
                    chunk_max = 800 
     | 
| 440 | 
         
             
                    overlap = 120
         
     | 
| 441 | 
         | 
| 442 | 
         
             
                chunks = chunk_text_by_chars(text_for_chunks, max_chars=chunk_max, overlap=overlap)
         
     | 
| 
         | 
|
| 443 | 
         | 
| 444 | 
         
            -
                #  
     | 
| 445 | 
         
            -
                chunk_target_sentences = 1 if length_choice == "short" else 2
         
     | 
| 446 | 
         
            -
                t_chunks_start = time.time()
         
     | 
| 447 | 
         
             
                try:
         
     | 
| 448 | 
         
            -
                    chunk_summaries = summarize_chunks_parallel(summarizer_pipe, chunks,  
     | 
| 449 | 
         
             
                except Exception as e:
         
     | 
| 450 | 
         
             
                    logger.exception("Chunk summarization orchestration failed: %s", e)
         
     | 
| 451 | 
         
            -
                    # fallback: simple extractive split
         
     | 
| 452 | 
         
             
                    chunk_summaries = [extractive_prefilter(c, top_k=3) for c in chunks]
         
     | 
| 453 | 
         
            -
                t_chunks_end = time.time()
         
     | 
| 454 | 
         | 
| 455 | 
         
            -
                #  
     | 
| 456 | 
         
             
                refine_pipe = _SUMMARIZER_CACHE.get("pegasus") or summarizer_pipe
         
     | 
| 457 | 
         
            -
                 
     | 
| 458 | 
         
            -
                final = refine_combined(refine_pipe, chunk_summaries, tone_choice,  
     | 
| 459 | 
         | 
| 460 | 
         
            -
                #  
     | 
| 461 | 
         
             
                if tone_choice == "bullet":
         
     | 
| 462 | 
         
             
                    parts = re.split(r'[\n\r]+|(?:\.\s+)|(?:;\s+)', final)
         
     | 
| 463 | 
         
             
                    bullets = [f"- {p.strip().rstrip('.')}" for p in parts if p.strip()]
         
     | 
| 464 | 
         
             
                    final = "\n".join(bullets[:20])
         
     | 
| 465 | 
         | 
| 466 | 
         
            -
                elapsed = time.time() -  
     | 
| 467 | 
         
             
                meta = {
         
     | 
| 468 | 
         
             
                    "length_choice": length_choice,
         
     | 
| 469 | 
         
             
                    "tone": tone_choice,
         
     | 
| 470 | 
         
            -
                    "model_requested": user_model_pref,
         
     | 
| 471 | 
         
             
                    "model_used": model_key,
         
     | 
| 472 | 
         
             
                    "chunks": len(chunks),
         
     | 
| 473 | 
         
             
                    "input_words": words_len,
         
     | 
| 474 | 
         
             
                    "time_seconds": round(elapsed, 2),
         
     | 
| 475 | 
         
            -
                    " 
     | 
| 476 | 
         
            -
                    "device": "cpu",
         
     | 
| 477 | 
         
            -
                    "workers_threads": MAX_WORKERS,
         
     | 
| 478 | 
         
            -
                    "per_chunk_timeout": CHUNK_TIMEOUT_SECONDS,
         
     | 
| 479 | 
         
            -
                    "refine_timeout": REFINE_TIMEOUT_SECONDS
         
     | 
| 480 | 
         
             
                }
         
     | 
| 481 | 
         
            -
             
     | 
| 482 | 
         
             
                return jsonify({"summary": final, "meta": meta})
         
     | 
| 483 | 
         | 
| 484 | 
         
             
            # -------------------------
         
     | 
| 485 | 
         
            -
            #  
     | 
| 486 | 
         
             
            # -------------------------
         
     | 
| 487 | 
         
             
            if __name__ == "__main__":
         
     | 
| 488 | 
         
            -
                # For  
     | 
| 489 | 
         
             
                app.run(host="0.0.0.0", port=7860, debug=False)
         
     | 
| 
         | 
|
| 1 | 
         
            +
            # main.py (REPLACE your existing file with this)
         
     | 
| 2 | 
         
             
            import os
         
     | 
| 3 | 
         
             
            os.environ['HF_HOME'] = '/tmp'
         
     | 
| 4 | 
         | 
| 
         | 
|
| 22 | 
         
             
            logger = logging.getLogger("summarizer")
         
     | 
| 23 | 
         | 
| 24 | 
         
             
            # -------------------------
         
     | 
| 25 | 
         
            +
            # Device selection (CPU by default)
         
     | 
| 26 | 
         
             
            # -------------------------
         
     | 
| 
         | 
|
| 27 | 
         
             
            USE_GPU = False
         
     | 
| 28 | 
         
             
            DEVICE = -1
         
     | 
| 29 | 
         
            +
            logger.info("Startup: forcing CPU usage for models (DEVICE=%s)", DEVICE)
         
     | 
| 30 | 
         | 
| 31 | 
         
             
            # -------------------------
         
     | 
| 32 | 
         
             
            # Model names and caches
         
     | 
| 33 | 
         
             
            # -------------------------
         
     | 
| 34 | 
         
             
            PEGASUS_MODEL = "google/pegasus-large"
         
     | 
| 35 | 
         
             
            LED_MODEL = "allenai/led-large-16384"
         
     | 
| 36 | 
         
            +
            DISTILBART_MODEL = "sshleifer/distilbart-cnn-12-6"
         
     | 
| 37 | 
         
            +
            PARAM_MODEL = "google/flan-t5-small"
         
     | 
| 38 | 
         | 
| 39 | 
         
             
            _SUMMARIZER_CACHE: Dict[str, Any] = {}
         
     | 
| 40 | 
         
             
            _PARAM_GENERATOR = None
         
     | 
| 41 | 
         
             
            _PREFERRED_SUMMARIZER_KEY: Optional[str] = None
         
     | 
| 42 | 
         | 
| 43 | 
         
             
            # -------------------------
         
     | 
| 44 | 
         
            +
            # Utilities: chunking, extractive fallback
         
     | 
| 45 | 
         
             
            # -------------------------
         
     | 
| 46 | 
         
             
            _STOPWORDS = {
         
     | 
| 47 | 
         
             
                "the","and","is","in","to","of","a","that","it","on","for","as","are","with","was","be","by","this","an","or","from","at","which","we","has","have"
         
     | 
| 
         | 
|
| 67 | 
         
             
                return " ".join(chosen)
         
     | 
| 68 | 
         | 
| 69 | 
         
             
            def chunk_text_by_chars(text: str, max_chars: int = 800, overlap: int = 120) -> List[str]:
         
     | 
| 
         | 
|
| 70 | 
         
             
                n = len(text)
         
     | 
| 71 | 
         
             
                if n <= max_chars:
         
     | 
| 72 | 
         
             
                    return [text]
         
     | 
| 
         | 
|
| 87 | 
         
             
                        start = end
         
     | 
| 88 | 
         
             
                return [p for p in parts if p]
         
     | 
| 89 | 
         | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 90 | 
         
             
            # -------------------------
         
     | 
| 91 | 
         
            +
            # safe loader (defined before any calls)
         
     | 
| 92 | 
         
             
            # -------------------------
         
     | 
| 93 | 
         
            +
            def safe_load_pipeline(model_name: str):
         
     | 
| 94 | 
         
            +
                """
         
     | 
| 95 | 
         
            +
                Try to load a summarization pipeline robustly:
         
     | 
| 96 | 
         
            +
                - try fast tokenizer first
         
     | 
| 97 | 
         
            +
                - if that fails, try use_fast=False
         
     | 
| 98 | 
         
            +
                - return pipeline or None if both fail
         
     | 
| 99 | 
         
            +
                """
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 100 | 
         
             
                try:
         
     | 
| 101 | 
         
            +
                    logger.info("Loading tokenizer/model for %s (fast)...", model_name)
         
     | 
| 102 | 
         
            +
                    tok = AutoTokenizer.from_pretrained(model_name)
         
     | 
| 103 | 
         
            +
                    model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
         
     | 
| 104 | 
         
            +
                    pipe = pipeline("summarization", model=model, tokenizer=tok, device=DEVICE)
         
     | 
| 105 | 
         
            +
                    logger.info("Loaded %s (fast tokenizer)", model_name)
         
     | 
| 106 | 
         
            +
                    return pipe
         
     | 
| 107 | 
         
            +
                except Exception as e_fast:
         
     | 
| 108 | 
         
            +
                    logger.warning("Fast tokenizer failed for %s: %s. Trying slow tokenizer...", model_name, e_fast)
         
     | 
| 109 | 
         
            +
                    try:
         
     | 
| 110 | 
         
            +
                        tok = AutoTokenizer.from_pretrained(model_name, use_fast=False)
         
     | 
| 111 | 
         
            +
                        model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
         
     | 
| 112 | 
         
            +
                        pipe = pipeline("summarization", model=model, tokenizer=tok, device=DEVICE)
         
     | 
| 113 | 
         
            +
                        logger.info("Loaded %s (slow tokenizer)", model_name)
         
     | 
| 114 | 
         
            +
                        return pipe
         
     | 
| 115 | 
         
            +
                    except Exception as e_slow:
         
     | 
| 116 | 
         
            +
                        logger.exception("Slow tokenizer failed for %s: %s", model_name, e_slow)
         
     | 
| 117 | 
         
            +
                        return None
         
     | 
| 118 | 
         | 
| 119 | 
         
             
            # -------------------------
         
     | 
| 120 | 
         
            +
            # get_summarizer: lazy load + cache + fallback
         
     | 
| 121 | 
         
             
            # -------------------------
         
     | 
| 122 | 
         
            +
            def get_summarizer(key: str):
         
     | 
| 123 | 
         
             
                """
         
     | 
| 124 | 
         
            +
                key: 'pegasus'|'led'|'distilbart'|'auto'
         
     | 
| 125 | 
         
            +
                returns a pipeline (cached), or raises RuntimeError if no pipeline can be loaded.
         
     | 
| 126 | 
         
             
                """
         
     | 
| 127 | 
         
            +
                key = (key or "auto").lower()
         
     | 
| 128 | 
         
            +
                if key == "auto":
         
     | 
| 129 | 
         
            +
                    key = _PREFERRED_SUMMARIZER_KEY or "distilbart"
         
     | 
| 130 | 
         
            +
             
     | 
| 131 | 
         
            +
                # direct mapping
         
     | 
| 132 | 
         
            +
                model_name = {
         
     | 
| 133 | 
         
            +
                    "pegasus": PEGASUS_MODEL,
         
     | 
| 134 | 
         
            +
                    "led": LED_MODEL,
         
     | 
| 135 | 
         
            +
                    "distilbart": DISTILBART_MODEL
         
     | 
| 136 | 
         
            +
                }.get(key, DISTILBART_MODEL)
         
     | 
| 137 | 
         
            +
             
     | 
| 138 | 
         
            +
                if key in _SUMMARIZER_CACHE:
         
     | 
| 139 | 
         
            +
                    return _SUMMARIZER_CACHE[key]
         
     | 
| 140 | 
         
            +
             
     | 
| 141 | 
         
            +
                # try to load
         
     | 
| 142 | 
         
            +
                logger.info("Attempting to lazy-load summarizer '%s' -> %s", key, model_name)
         
     | 
| 143 | 
         
            +
                pipe = safe_load_pipeline(model_name)
         
     | 
| 144 | 
         
            +
                if pipe:
         
     | 
| 145 | 
         
            +
                    _SUMMARIZER_CACHE[key] = pipe
         
     | 
| 146 | 
         
            +
                    return pipe
         
     | 
| 147 | 
         
            +
             
     | 
| 148 | 
         
            +
                # fallback attempts
         
     | 
| 149 | 
         
            +
                logger.warning("Failed to load %s. Trying distilbart fallback.", key)
         
     | 
| 150 | 
         
            +
                if "distilbart" in _SUMMARIZER_CACHE:
         
     | 
| 151 | 
         
            +
                    return _SUMMARIZER_CACHE["distilbart"]
         
     | 
| 152 | 
         
            +
                fb = safe_load_pipeline(DISTILBART_MODEL)
         
     | 
| 153 | 
         
            +
                if fb:
         
     | 
| 154 | 
         
            +
                    _SUMMARIZER_CACHE["distilbart"] = fb
         
     | 
| 155 | 
         
            +
                    return fb
         
     | 
| 156 | 
         
            +
             
     | 
| 157 | 
         
            +
                # nothing works
         
     | 
| 158 | 
         
            +
                raise RuntimeError("No summarizer available. Install required libraries and/or choose smaller model.")
         
     | 
| 159 | 
         
            +
             
     | 
| 160 | 
         
            +
            # -------------------------
         
     | 
| 161 | 
         
            +
            # Generation strategy + small helpers
         
     | 
| 162 | 
         
            +
            # -------------------------
         
     | 
| 163 | 
         
            +
            def summarize_with_model(pipe, text_prompt: str, short_target: bool = False) -> str:
         
     | 
| 164 | 
         
             
                model_name = getattr(pipe.model.config, "name_or_path", "") or ""
         
     | 
| 165 | 
         
            +
                is_led = "led" in model_name.lower() or "longformer" in model_name.lower()
         
     | 
| 
         | 
|
| 166 | 
         | 
| 167 | 
         
            +
                # Fast sampling pass
         
     | 
| 168 | 
         
             
                fast_cfg = {
         
     | 
| 169 | 
         
             
                    "max_new_tokens": 64 if short_target else (120 if not is_led else 240),
         
     | 
| 170 | 
         
             
                    "do_sample": True,
         
     | 
| 171 | 
         
            +
                    "top_p": 0.92,
         
     | 
| 172 | 
         
             
                    "temperature": 0.85,
         
     | 
| 173 | 
         
             
                    "num_beams": 1,
         
     | 
| 174 | 
         
             
                    "early_stopping": True,
         
     | 
| 175 | 
         
             
                    "no_repeat_ngram_size": 3,
         
     | 
| 176 | 
         
             
                }
         
     | 
| 
         | 
|
| 177 | 
         
             
                try:
         
     | 
| 178 | 
         
            +
                    return pipe(text_prompt, **fast_cfg)[0].get("summary_text","").strip()
         
     | 
| 
         | 
|
| 179 | 
         
             
                except Exception as e:
         
     | 
| 180 | 
         
            +
                    logger.warning("Fast pass failed: %s, trying quality pass...", e)
         
     | 
| 181 | 
         | 
| 
         | 
|
| 182 | 
         
             
                quality_cfg = {
         
     | 
| 183 | 
         
             
                    "max_new_tokens": 140 if not is_led else 320,
         
     | 
| 184 | 
         
             
                    "do_sample": False,
         
     | 
| 
         | 
|
| 186 | 
         
             
                    "early_stopping": True,
         
     | 
| 187 | 
         
             
                    "no_repeat_ngram_size": 3,
         
     | 
| 188 | 
         
             
                }
         
     | 
| 
         | 
|
| 189 | 
         
             
                try:
         
     | 
| 190 | 
         
            +
                    return pipe(text_prompt, **quality_cfg)[0].get("summary_text","").strip()
         
     | 
| 191 | 
         
            +
                except Exception as e:
         
     | 
| 192 | 
         
            +
                    logger.exception("Quality pass failed: %s", e)
         
     | 
| 
         | 
|
| 193 | 
         | 
| 194 | 
         
            +
                # fallback extractive
         
     | 
| 195 | 
         
             
                try:
         
     | 
| 196 | 
         
             
                    return extractive_prefilter(text_prompt, top_k=3)
         
     | 
| 197 | 
         
             
                except Exception:
         
     | 
| 198 | 
         
            +
                    return "Summarization failed; try shorter input."
         
     | 
| 199 | 
         | 
| 200 | 
         
             
            # -------------------------
         
     | 
| 201 | 
         
            +
            # Param generator (AI decision) - lazy loader
         
     | 
| 202 | 
         
             
            # -------------------------
         
     | 
| 203 | 
         
            +
            def get_param_generator():
         
     | 
| 204 | 
         
            +
                global _PARAM_GENERATOR
         
     | 
| 205 | 
         
            +
                if _PARAM_GENERATOR is not None:
         
     | 
| 206 | 
         
            +
                    return _PARAM_GENERATOR
         
     | 
| 207 | 
         
            +
                # try to load text2text pipeline for PARAM_MODEL
         
     | 
| 208 | 
         
            +
                try:
         
     | 
| 209 | 
         
            +
                    logger.info("Loading param-generator (text2text) lazily: %s", PARAM_MODEL)
         
     | 
| 210 | 
         
            +
                    tok = AutoTokenizer.from_pretrained(PARAM_MODEL)
         
     | 
| 211 | 
         
            +
                    mod = AutoModelForSeq2SeqLM.from_pretrained(PARAM_MODEL)
         
     | 
| 212 | 
         
            +
                    _PARAM_GENERATOR = pipeline("text2text-generation", model=mod, tokenizer=tok, device=DEVICE)
         
     | 
| 213 | 
         
            +
                    logger.info("Param-generator loaded.")
         
     | 
| 214 | 
         
            +
                    return _PARAM_GENERATOR
         
     | 
| 215 | 
         
            +
                except Exception as e:
         
     | 
| 216 | 
         
            +
                    logger.exception("Param-generator lazy load failed: %s", e)
         
     | 
| 217 | 
         
            +
                    _PARAM_GENERATOR = None
         
     | 
| 218 | 
         
            +
                    return None
         
     | 
| 219 | 
         
            +
             
     | 
| 220 | 
         
             
            def generate_summarization_config(text: str) -> Dict[str, Any]:
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 221 | 
         
             
                defaults = {"short": (12, 50), "medium": (50, 130), "long": (130, 300)}
         
     | 
| 222 | 
         
            +
                pg = get_param_generator()
         
     | 
| 223 | 
         
            +
                if pg is None:
         
     | 
| 224 | 
         
             
                    words = len(text.split())
         
     | 
| 225 | 
         
             
                    length = "short" if words < 150 else ("medium" if words < 800 else "long")
         
     | 
| 226 | 
         
             
                    mn, mx = defaults[length]
         
     | 
| 227 | 
         
             
                    return {"length": length, "min_length": mn, "max_length": mx, "tone": "neutral"}
         
     | 
| 228 | 
         | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 229 | 
         
             
                prompt = (
         
     | 
| 230 | 
         
            +
                    "Recommend summarization settings for this text. Answer ONLY with JSON like:\n"
         
     | 
| 231 | 
         
             
                    '{"length":"short|medium|long","tone":"neutral|formal|casual|bullet","min_words":MIN,"max_words":MAX}\n\n'
         
     | 
| 232 | 
         
             
                    "Text:\n'''"
         
     | 
| 233 | 
         
             
                    + text[:3000] + "'''"
         
     | 
| 234 | 
         
             
                )
         
     | 
| 
         | 
|
| 235 | 
         
             
                try:
         
     | 
| 236 | 
         
            +
                    out_item = pg(prompt, max_new_tokens=64, do_sample=False, num_beams=1)[0]
         
     | 
| 237 | 
         
            +
                    out = out_item.get("generated_text") or out_item.get("summary_text") or ""
         
     | 
| 
         | 
|
| 238 | 
         
             
                    out = (out or "").strip()
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 239 | 
         
             
                    if not out:
         
     | 
| 240 | 
         
             
                        raise ValueError("Empty param-generator output")
         
     | 
| 241 | 
         
            +
                    # reject noisy echo outputs
         
     | 
| 242 | 
         
             
                    input_words = set(w.lower() for w in re.findall(r"\w+", text)[:200])
         
     | 
| 243 | 
         
             
                    out_words = set(w.lower() for w in re.findall(r"\w+", out)[:200])
         
     | 
| 244 | 
         
            +
                    if len(input_words) and (len(input_words & out_words) / max(1, len(input_words))) > 0.4:
         
     | 
| 245 | 
         
            +
                        logger.warning("Param-generator appears to echo input; using heuristic fallback.")
         
     | 
| 246 | 
         
            +
                        words = len(text.split())
         
     | 
| 247 | 
         
            +
                        length = "short" if words < 150 else ("medium" if words < 800 else "long")
         
     | 
| 248 | 
         
            +
                        mn, mx = defaults[length]
         
     | 
| 249 | 
         
            +
                        return {"length": length, "min_length": mn, "max_length": mx, "tone": "neutral"}
         
     | 
| 250 | 
         | 
| 
         | 
|
| 251 | 
         
             
                    jmatch = re.search(r"\{.*\}", out, re.DOTALL)
         
     | 
| 252 | 
         
             
                    if jmatch:
         
     | 
| 253 | 
         
             
                        raw = jmatch.group().replace("'", '"')
         
     | 
| 254 | 
         
             
                        cfg = json.loads(raw)
         
     | 
| 255 | 
         
             
                    else:
         
     | 
| 
         | 
|
| 256 | 
         
             
                        cfg = None
         
     | 
| 257 | 
         | 
| 258 | 
         
             
                    if not cfg or not isinstance(cfg, dict):
         
     | 
| 259 | 
         
            +
                        raise ValueError("Param output not parseable")
         
     | 
| 260 | 
         
            +
                    length = cfg.get("length","medium").lower()
         
     | 
| 261 | 
         
            +
                    tone = cfg.get("tone","neutral").lower()
         
     | 
| 262 | 
         
            +
                    mn = int(cfg.get("min_words") or cfg.get("min_length") or defaults[length][0])
         
     | 
| 263 | 
         
            +
                    mx = int(cfg.get("max_words") or cfg.get("max_length") or defaults[length][1])
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 264 | 
         
             
                    mn = max(5, min(mn, 2000))
         
     | 
| 265 | 
         
             
                    mx = max(mn + 5, min(mx, 4000))
         
     | 
| 
         | 
|
| 266 | 
         
             
                    return {"length": length, "min_length": mn, "max_length": mx, "tone": tone}
         
     | 
| 
         | 
|
| 267 | 
         
             
                except Exception as e:
         
     | 
| 268 | 
         
            +
                    logger.exception("Param-generator parse failed: %s", e)
         
     | 
| 269 | 
         
            +
                    words = len(text.split())
         
     | 
| 270 | 
         
            +
                    length = "short" if words < 150 else ("medium" if words < 800 else "long")
         
     | 
| 271 | 
         
            +
                    mn, mx = defaults[length]
         
     | 
| 272 | 
         
            +
                    return {"length": length, "min_length": mn, "max_length": mx, "tone": "neutral"}
         
     | 
| 273 | 
         | 
| 274 | 
         
             
            # -------------------------
         
     | 
| 275 | 
         
            +
            # Threaded chunk summarization with per-chunk timeout (to prevent hang)
         
     | 
| 276 | 
         
             
            # -------------------------
         
     | 
| 277 | 
         
            +
            executor = ThreadPoolExecutor(max_workers=min(8, max(2, (os.cpu_count() or 2))))
         
     | 
| 278 | 
         
            +
            CHUNK_TIMEOUT_SECONDS = 28
         
     | 
| 279 | 
         
            +
            REFINE_TIMEOUT_SECONDS = 60
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 280 | 
         | 
| 281 | 
         
             
            def summarize_chunks_parallel(pipe, chunks: List[str], chunk_target: int) -> List[str]:
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 282 | 
         
             
                futures = {}
         
     | 
| 283 | 
         
             
                results = [None] * len(chunks)
         
     | 
| 284 | 
         
             
                for idx, chunk in enumerate(chunks):
         
     | 
| 
         | 
|
| 287 | 
         
             
                    futures[fut] = idx
         
     | 
| 288 | 
         | 
| 289 | 
         
             
                start = time.time()
         
     | 
| 290 | 
         
            +
                for fut in as_completed(futures):
         
     | 
| 291 | 
         
             
                    idx = futures[fut]
         
     | 
| 292 | 
         
             
                    try:
         
     | 
| 293 | 
         
             
                        remaining = max(0.1, CHUNK_TIMEOUT_SECONDS - (time.time() - start))
         
     | 
| 
         | 
|
| 294 | 
         
             
                        results[idx] = fut.result(timeout=remaining)
         
     | 
| 295 | 
         
             
                    except TimeoutError:
         
     | 
| 296 | 
         
             
                        logger.warning("Chunk %d timed out; using extractive fallback.", idx)
         
     | 
| 297 | 
         
             
                        results[idx] = extractive_prefilter(chunks[idx], top_k=3)
         
     | 
| 298 | 
         
             
                    except Exception as e:
         
     | 
| 299 | 
         
            +
                        logger.exception("Chunk %d failed: %s; falling back", idx, e)
         
     | 
| 300 | 
         
             
                        results[idx] = extractive_prefilter(chunks[idx], top_k=3)
         
     | 
| 
         | 
|
| 301 | 
         
             
                for i, r in enumerate(results):
         
     | 
| 302 | 
         
             
                    if not r:
         
     | 
| 303 | 
         
             
                        results[i] = extractive_prefilter(chunks[i], top_k=3)
         
     | 
| 304 | 
         
             
                return results
         
     | 
| 305 | 
         | 
| 306 | 
         
             
            # -------------------------
         
     | 
| 307 | 
         
            +
            # Prompt helpers and refine
         
     | 
| 308 | 
         
             
            # -------------------------
         
     | 
| 309 | 
         
             
            def apply_tone_instruction(text: str, tone: str, target_sentences: Optional[int] = None) -> str:
         
     | 
| 310 | 
         
             
                tone = (tone or "neutral").lower()
         
     | 
| 
         | 
|
| 312 | 
         
             
                    instr = "Produce concise bullet points. Each bullet <= 20 words. No extra commentary."
         
     | 
| 313 | 
         
             
                elif tone == "short":
         
     | 
| 314 | 
         
             
                    ts = target_sentences or 1
         
     | 
| 315 | 
         
            +
                    instr = f"Summarize in {ts} sentence{'s' if ts>1 else ''}. Be abstractive."
         
     | 
| 316 | 
         
             
                elif tone == "formal":
         
     | 
| 317 | 
         
            +
                    instr = "Summarize in a formal, professional tone (2-4 sentences)."
         
     | 
| 318 | 
         
             
                elif tone == "casual":
         
     | 
| 319 | 
         
            +
                    instr = "Summarize in a casual, conversational tone (1-3 sentences)."
         
     | 
| 320 | 
         
             
                elif tone == "long":
         
     | 
| 321 | 
         
            +
                    instr = "Provide a structured summary (4-8 sentences)."
         
     | 
| 322 | 
         
             
                else:
         
     | 
| 323 | 
         
             
                    instr = "Summarize in 2-3 clear sentences."
         
     | 
| 324 | 
         
             
                instr += " Do not repeat information. Prefer rephrasing."
         
     | 
| 
         | 
|
| 329 | 
         
             
                if len(combined.split()) > 1200:
         
     | 
| 330 | 
         
             
                    combined = extractive_prefilter(combined, top_k=20)
         
     | 
| 331 | 
         
             
                prompt = apply_tone_instruction(combined, tone, target_sentences=final_target_sentences)
         
     | 
| 
         | 
|
| 332 | 
         
             
                fut = executor.submit(summarize_with_model, pipe, prompt, short_target=False)
         
     | 
| 333 | 
         
             
                try:
         
     | 
| 334 | 
         
             
                    return fut.result(timeout=REFINE_TIMEOUT_SECONDS)
         
     | 
| 335 | 
         
             
                except TimeoutError:
         
     | 
| 336 | 
         
            +
                    logger.warning("Refine timed out; returning concatenated chunk summaries.")
         
     | 
| 337 | 
         
             
                    return " ".join(summaries_list[:6])
         
     | 
| 338 | 
         
             
                except Exception as e:
         
     | 
| 339 | 
         
            +
                    logger.exception("Refine failed: %s", e)
         
     | 
| 340 | 
         
             
                    return " ".join(summaries_list[:6])
         
     | 
| 341 | 
         | 
| 342 | 
         
             
            # -------------------------
         
     | 
| 343 | 
         
            +
            # Routes
         
     | 
| 344 | 
         
             
            # -------------------------
         
     | 
| 345 | 
         
             
            @app.route("/", methods=["GET"])
         
     | 
| 346 | 
         
             
            def home():
         
     | 
| 347 | 
         
             
                try:
         
     | 
| 348 | 
         
             
                    return render_template("index.html")
         
     | 
| 349 | 
         
             
                except Exception:
         
     | 
| 350 | 
         
            +
                    return "Summarizer (lazy-load) — POST /summarize with JSON {text:'...'}", 200
         
     | 
| 351 | 
         
            +
             
     | 
| 352 | 
         
            +
            @app.route("/preload", methods=["POST"])
         
     | 
| 353 | 
         
            +
            def preload_models():
         
     | 
| 354 | 
         
            +
                """
         
     | 
| 355 | 
         
            +
                Explicit endpoint to attempt preloading heavy models.
         
     | 
| 356 | 
         
            +
                Call this only when you want the process to attempt loading Pegasus/LED (may be slow).
         
     | 
| 357 | 
         
            +
                """
         
     | 
| 358 | 
         
            +
                results = {}
         
     | 
| 359 | 
         
            +
                for key, model_name in [("pegasus", PEGASUS_MODEL), ("led", LED_MODEL), ("distilbart", DISTILBART_MODEL)]:
         
     | 
| 360 | 
         
            +
                    if key in _SUMMARIZER_CACHE:
         
     | 
| 361 | 
         
            +
                        results[key] = "already_loaded"
         
     | 
| 362 | 
         
            +
                        continue
         
     | 
| 363 | 
         
            +
                    try:
         
     | 
| 364 | 
         
            +
                        p = safe_load_pipeline(model_name)
         
     | 
| 365 | 
         
            +
                        if p:
         
     | 
| 366 | 
         
            +
                            _SUMMARIZER_CACHE[key] = p
         
     | 
| 367 | 
         
            +
                            results[key] = "loaded"
         
     | 
| 368 | 
         
            +
                        else:
         
     | 
| 369 | 
         
            +
                            results[key] = "failed"
         
     | 
| 370 | 
         
            +
                    except Exception as e:
         
     | 
| 371 | 
         
            +
                        results[key] = f"error: {e}"
         
     | 
| 372 | 
         
            +
                return jsonify(results)
         
     | 
| 373 | 
         | 
| 374 | 
         
             
            @app.route("/summarize", methods=["POST"])
         
     | 
| 375 | 
         
             
            def summarize_route():
         
     | 
| 376 | 
         
            +
                t0 = time.time()
         
     | 
| 377 | 
         
             
                data = request.get_json(force=True) or {}
         
     | 
| 378 | 
         
             
                text = (data.get("text") or "").strip()[:90000]
         
     | 
| 379 | 
         
             
                user_model_pref = (data.get("model") or "auto").lower()
         
     | 
| 
         | 
|
| 383 | 
         
             
                if not text or len(text.split()) < 5:
         
     | 
| 384 | 
         
             
                    return jsonify({"error": "Input too short."}), 400
         
     | 
| 385 | 
         | 
| 386 | 
         
            +
                # decide settings
         
     | 
| 387 | 
         
            +
                if requested_length in ("auto","ai") or requested_tone in ("auto","ai"):
         
     | 
| 388 | 
         
             
                    cfg = generate_summarization_config(text)
         
     | 
| 389 | 
         
            +
                    length_choice = cfg.get("length","medium")
         
     | 
| 390 | 
         
            +
                    tone_choice = cfg.get("tone","neutral")
         
     | 
| 391 | 
         
             
                else:
         
     | 
| 392 | 
         
             
                    length_choice = requested_length if requested_length in ("short","medium","long") else "medium"
         
     | 
| 393 | 
         
             
                    tone_choice = requested_tone if requested_tone in ("neutral","formal","casual","bullet") else "neutral"
         
     | 
| 394 | 
         | 
| 395 | 
         
            +
                # model selection logic
         
     | 
| 396 | 
         
             
                words_len = len(text.split())
         
     | 
| 397 | 
         
             
                prefer_led = False
         
     | 
| 398 | 
         
             
                if user_model_pref == "led":
         
     | 
| 
         | 
|
| 402 | 
         
             
                else:
         
     | 
| 403 | 
         
             
                    if length_choice == "long" or words_len > 3000:
         
     | 
| 404 | 
         
             
                        prefer_led = True
         
     | 
| 
         | 
|
| 
         | 
|
| 405 | 
         | 
| 
         | 
|
| 406 | 
         
             
                model_key = "led" if prefer_led else (_PREFERRED_SUMMARIZER_KEY or "distilbart")
         
     | 
| 407 | 
         
            +
                try:
         
     | 
| 408 | 
         
            +
                    summarizer_pipe = get_summarizer(model_key)
         
     | 
| 409 | 
         
            +
                except Exception as e:
         
     | 
| 410 | 
         
            +
                    logger.exception("get_summarizer failed (%s). Falling back to distilbart.", e)
         
     | 
| 411 | 
         
            +
                    summarizer_pipe = get_summarizer("distilbart")
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 412 | 
         
             
                    model_key = "distilbart"
         
     | 
| 
         | 
|
| 413 | 
         | 
| 414 | 
         
            +
                # prefilter very long inputs for non-LED
         
     | 
| 415 | 
         
             
                if model_key != "led" and words_len > 2500:
         
     | 
| 416 | 
         
             
                    text_for_chunks = extractive_prefilter(text, top_k=40)
         
     | 
| 417 | 
         
             
                else:
         
     | 
| 418 | 
         
             
                    text_for_chunks = text
         
     | 
| 419 | 
         | 
| 420 | 
         
            +
                # chunk sizing
         
     | 
| 421 | 
         
             
                if model_key == "led":
         
     | 
| 422 | 
         
             
                    chunk_max = 6000
         
     | 
| 423 | 
         
             
                    overlap = 400
         
     | 
| 424 | 
         
             
                else:
         
     | 
| 425 | 
         
            +
                    chunk_max = 800
         
     | 
| 426 | 
         
             
                    overlap = 120
         
     | 
| 427 | 
         | 
| 428 | 
         
             
                chunks = chunk_text_by_chars(text_for_chunks, max_chars=chunk_max, overlap=overlap)
         
     | 
| 429 | 
         
            +
                chunk_target = 1 if length_choice == "short" else 2
         
     | 
| 430 | 
         | 
| 431 | 
         
            +
                # summarize chunks in parallel
         
     | 
| 
         | 
|
| 
         | 
|
| 432 | 
         
             
                try:
         
     | 
| 433 | 
         
            +
                    chunk_summaries = summarize_chunks_parallel(summarizer_pipe, chunks, chunk_target)
         
     | 
| 434 | 
         
             
                except Exception as e:
         
     | 
| 435 | 
         
             
                    logger.exception("Chunk summarization orchestration failed: %s", e)
         
     | 
| 
         | 
|
| 436 | 
         
             
                    chunk_summaries = [extractive_prefilter(c, top_k=3) for c in chunks]
         
     | 
| 
         | 
|
| 437 | 
         | 
| 438 | 
         
            +
                # refine step — prefer Pegasus if loaded, otherwise use current pipe
         
     | 
| 439 | 
         
             
                refine_pipe = _SUMMARIZER_CACHE.get("pegasus") or summarizer_pipe
         
     | 
| 440 | 
         
            +
                final_target_sentences = {"short":1,"medium":3,"long":6}.get(length_choice, 3)
         
     | 
| 441 | 
         
            +
                final = refine_combined(refine_pipe, chunk_summaries, tone_choice, final_target_sentences)
         
     | 
| 442 | 
         | 
| 443 | 
         
            +
                # bullet postprocess
         
     | 
| 444 | 
         
             
                if tone_choice == "bullet":
         
     | 
| 445 | 
         
             
                    parts = re.split(r'[\n\r]+|(?:\.\s+)|(?:;\s+)', final)
         
     | 
| 446 | 
         
             
                    bullets = [f"- {p.strip().rstrip('.')}" for p in parts if p.strip()]
         
     | 
| 447 | 
         
             
                    final = "\n".join(bullets[:20])
         
     | 
| 448 | 
         | 
| 449 | 
         
            +
                elapsed = time.time() - t0
         
     | 
| 450 | 
         
             
                meta = {
         
     | 
| 451 | 
         
             
                    "length_choice": length_choice,
         
     | 
| 452 | 
         
             
                    "tone": tone_choice,
         
     | 
| 
         | 
|
| 453 | 
         
             
                    "model_used": model_key,
         
     | 
| 454 | 
         
             
                    "chunks": len(chunks),
         
     | 
| 455 | 
         
             
                    "input_words": words_len,
         
     | 
| 456 | 
         
             
                    "time_seconds": round(elapsed, 2),
         
     | 
| 457 | 
         
            +
                    "device": "cpu"
         
     | 
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 
         | 
|
| 458 | 
         
             
                }
         
     | 
| 
         | 
|
| 459 | 
         
             
                return jsonify({"summary": final, "meta": meta})
         
     | 
| 460 | 
         | 
| 461 | 
         
             
            # -------------------------
         
     | 
| 462 | 
         
            +
            # Local run (safe)
         
     | 
| 463 | 
         
             
            # -------------------------
         
     | 
| 464 | 
         
             
            if __name__ == "__main__":
         
     | 
| 465 | 
         
            +
                # For local testing you may call preload_models_at_startup manually or use /preload.
         
     | 
| 466 | 
         
             
                app.run(host="0.0.0.0", port=7860, debug=False)
         
     |