Update src/ai_processor.py

src/ai_processor.py

@@ -1,73 +1,42 @@
 # smartheal_ai_processor.py
-# Verbose, instrumented version — preserves public class/function names
-# Turn on deep logging: export LOGLEVEL=DEBUG SMARTHEAL_DEBUG=1
+# Full, functional module with conditional Spaces GPU support and CPU fallbacks.
 
 import os
 import time
 import logging
 from datetime import datetime
-from typing import Optional, Dict, List, Tuple
-
-# ---- Environment defaults ----
-os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
-os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
-LOGLEVEL = os.getenv("LOGLEVEL", "INFO").upper()
-SMARTHEAL_DEBUG = os.getenv("SMARTHEAL_DEBUG", "0") == "1"
+from typing import Optional, Dict, List
 
 import cv2
 import numpy as np
 from PIL import Image
-from PIL.ExifTags import TAGS
-
-# --- Logging config ---
-logging.basicConfig(
-    level=getattr(logging, LOGLEVEL, logging.INFO),
-    format="%(asctime)s - %(levelname)s - %(message)s",
-)
-
-def _log_kv(prefix: str, kv: Dict):
-    logging.debug(prefix + " | " + " | ".join(f"{k}={v}" for k, v in kv.items()))
-
-# --- Optional Spaces GPU stub (harmless) ---
-try:
-    import spaces as _spaces
-    @_spaces.GPU(enable_queue=False)
-    def smartheal_gpu_stub(ping: int = 0) -> str:
-        return "ready"
-    logging.info("Registered @spaces.GPU stub (enable_queue=False).")
-except Exception:
-    pass
 
+# =============== LOGGING SETUP ===============
+logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
+
+# =============== CONFIGURATION ===============
 UPLOADS_DIR = "uploads"
 os.makedirs(UPLOADS_DIR, exist_ok=True)
 
 HF_TOKEN = os.getenv("HF_TOKEN", None)
 YOLO_MODEL_PATH = "src/best.pt"
-SEG_MODEL_PATH = "src/segmentation_model.h5"   # optional
+SEG_MODEL_PATH = "src/segmentation_model.h5"  # optional
 GUIDELINE_PDFS = ["src/eHealth in Wound Care.pdf", "src/IWGDF Guideline.pdf", "src/evaluation.pdf"]
-DATASET_ID = "SmartHeal/wound-image-uploads"
-DEFAULT_PX_PER_CM = 38.0
-PX_PER_CM_MIN, PX_PER_CM_MAX = 5.0, 1200.0
-
-# Segmentation preprocessing knobs
-SEG_EXPECTS_RGB = os.getenv("SEG_EXPECTS_RGB", "1") == "1"  # most TF models trained on RGB
-SEG_NORM = os.getenv("SEG_NORM", "0to1")                    # "0to1" | "imagenet"
-SEG_THRESH = float(os.getenv("SEG_THRESH", "0.5"))
+DATASET_ID = "SmartHeal/wound-image-uploads"  # optional (set HF_TOKEN too)
+PIXELS_PER_CM = 38  # heuristic
 
+# =============== GLOBAL CACHES ===============
 models_cache: Dict[str, object] = {}
 knowledge_base_cache: Dict[str, object] = {}
 
-# ---------- Lazy imports ----------
+# ---------- Optional imports guarded ----------
 def _import_ultralytics():
     from ultralytics import YOLO
     return YOLO
 
 def _import_tf_loader():
     import tensorflow as tf
-    try:
-        tf.config.set_visible_devices([], "GPU")  # keep TF on CPU
-    except Exception:
-        pass
+    tf.config.set_visible_devices([], "GPU")  # force CPU
     from tensorflow.keras.models import load_model
     return load_model
 
@@ -91,50 +60,107 @@ def _import_hf_hub():
     from huggingface_hub import HfApi, HfFolder
     return HfApi, HfFolder
 
-# ---------- VLM (disabled by default) ----------
-def generate_medgemma_report(
-    patient_info: str,
-    visual_results: Dict,
-    guideline_context: str,
-    image_pil: Image.Image,
-    max_new_tokens: Optional[int] = None,
-) -> str:
-    if os.getenv("SMARTHEAL_ENABLE_VLM", "0") != "1":
-        return "⚠️ VLM disabled"
+# =============== SPACES GPU CONDITIONAL ===============
+def _spaces_gpu_available() -> bool:
     try:
+        import torch
+        return bool(torch.cuda.is_available())
+    except Exception:
+        return False
+
+def _spaces_lib_available() -> bool:
+    try:
+        import spaces  # noqa
+        return True
+    except Exception:
+        return False
+
+HAVE_SPACES_GPU = _spaces_gpu_available() and _spaces_lib_available()
+
+if HAVE_SPACES_GPU:
+    import spaces  # define only if available & GPU present
+
+    @spaces.GPU(enable_queue=True, duration=90)
+    def generate_medgemma_report_with_timeout(
+        patient_info: str,
+        visual_results: Dict,
+        guideline_context: str,
+        image_pil: Image.Image,
+        max_new_tokens: Optional[int] = None,
+    ) -> str:
+        """Runs on Spaces GPU only; callers keep one signature on both paths."""
+        import torch
         from transformers import pipeline
-        pipe = pipeline(
-            task="image-text-to-text",
-            model="google/medgemma-4b-it",
-            device_map=None,
-            token=HF_TOKEN,
-            trust_remote_code=True,
-            model_kwargs={"low_cpu_mem_usage": True},
-        )
-        prompt = (
-            "You are a medical AI assistant. Analyze this wound image and patient data.\n\n"
-            f"Patient: {patient_info}\n"
-            f"Wound: {visual_results.get('wound_type', 'Unknown')} - "
-            f"{visual_results.get('length_cm', 0)}×{visual_results.get('breadth_cm', 0)} cm\n\n"
-            "Provide a structured report with:\n"
-            "1. Clinical Summary\n2. Treatment Recommendations\n3. Risk Assessment\n4. Monitoring Plan\n"
-        )
-        messages = [{"role": "user", "content": [
-            {"type": "image", "image": image_pil},
-            {"type": "text",  "text": prompt},
-        ]}]
-        out = pipe(text=messages, max_new_tokens=max_new_tokens or 600, do_sample=False, temperature=0.7)
-        if out and len(out) > 0:
+        try:
+            torch.cuda.empty_cache()
+
+            prompt = f"""
+You are a medical AI assistant. Analyze this wound image and patient data.
+Patient: {patient_info}
+Wound: {visual_results.get('wound_type', 'Unknown')} - {visual_results.get('length_cm', 0)}×{visual_results.get('breadth_cm', 0)} cm
+Provide a structured report with:
+1. Clinical Summary
+2. Treatment Recommendations
+3. Risk Assessment
+4. Monitoring Plan
+""".strip()
+
+            pipe = pipeline(
+                "image-text-to-text",
+                model="google/medgemma-4b-it",
+                torch_dtype=torch.bfloat16,
+                device_map="auto",
+                token=HF_TOKEN,
+                model_kwargs={"low_cpu_mem_usage": True, "use_cache": True},
+            )
+
+            messages = [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image", "image": image_pil},
+                        {"type": "text", "text": prompt},
+                    ],
+                }
+            ]
+
+            t0 = time.time()
+            out = pipe(
+                text=messages,
+                max_new_tokens=max_new_tokens or 800,
+                do_sample=False,
+                temperature=0.7,
+                pad_token_id=pipe.tokenizer.eos_token_id,
+            )
+            logging.info(f"✅ MedGemma completed in {time.time() - t0:.2f}s")
+
+            if out and len(out) > 0:
+                # Defensive extraction
+                try:
+                    return out[0]["generated_text"][-1].get("content", "").strip() or "⚠️ Empty response"
+                except Exception:
+                    return (out[0].get("generated_text", "") or "").strip() or "⚠️ Empty response"
+            return "⚠️ No output generated"
+        except Exception as e:
+            logging.error(f"❌ MedGemma generation error: {e}")
+            return f"❌ Report generation failed: {str(e)}"
+        finally:
             try:
-                return out[0]["generated_text"][-1].get("content", "").strip() or "⚠️ Empty response"
+                torch.cuda.empty_cache()
             except Exception:
-                return (out[0].get("generated_text", "") or "").strip() or "⚠️ Empty response"
-        return "⚠️ No output generated"
-    except Exception as e:
-        logging.error(f"❌ MedGemma generation error: {e}")
-        return "⚠️ VLM error"
+                pass
+else:
+    def generate_medgemma_report_with_timeout(
+        patient_info: str,
+        visual_results: Dict,
+        guideline_context: str,
+        image_pil: Image.Image,
+        max_new_tokens: Optional[int] = None,
+    ) -> str:
+        """CPU-only path: return a warning so caller uses fallback."""
+        return "⚠️ GPU not available"
 
-# ---------- Initialize CPU models ----------
+# =============== MODEL INITIALIZATION (CPU-SAFE) ===============
 def load_yolo_model():
     YOLO = _import_ultralytics()
     return YOLO(YOLO_MODEL_PATH)
@@ -145,25 +171,32 @@ def load_segmentation_model():
 
 def load_classification_pipeline():
     pipe = _import_hf_cls()
-    return pipe("image-classification", model="Hemg/Wound-classification", token=HF_TOKEN, device="cpu")
+    return pipe(
+        "image-classification",
+        model="Hemg/Wound-classification",
+        token=HF_TOKEN,
+        device="cpu",
+    )
 
 def load_embedding_model():
     Emb = _import_embeddings()
     return Emb(model_name="sentence-transformers/all-MiniLM-L6-v2", model_kwargs={"device": "cpu"})
 
 def initialize_cpu_models() -> None:
+    """Initialize all CPU-only models once with robust fallbacks."""
+    # Hugging Face auth (optional)
     if HF_TOKEN:
         try:
             HfApi, HfFolder = _import_hf_hub()
             HfFolder.save_token(HF_TOKEN)
-            logging.info("✅ HF token set")
+            logging.info("✅ HuggingFace token set")
         except Exception as e:
             logging.warning(f"HF token save failed: {e}")
 
     if "det" not in models_cache:
         try:
             models_cache["det"] = load_yolo_model()
-            logging.info("✅ YOLO loaded (CPU)")
+            logging.info("✅ YOLO model loaded (CPU)")
         except Exception as e:
             logging.error(f"YOLO load failed: {e}")
 
@@ -171,46 +204,46 @@ def initialize_cpu_models() -> None:
         try:
             if os.path.exists(SEG_MODEL_PATH):
                 models_cache["seg"] = load_segmentation_model()
-                m = models_cache["seg"]
-                ishape = getattr(m, "input_shape", None)
-                oshape = getattr(m, "output_shape", None)
-                logging.info(f"✅ Segmentation model loaded (CPU) | input_shape={ishape} output_shape={oshape}")
+                logging.info("✅ Segmentation model loaded (CPU)")
             else:
                 models_cache["seg"] = None
-                logging.warning("Segmentation model file missing; skipping.")
+                logging.warning("Segmentation model file not found; skipping seg.")
         except Exception as e:
             models_cache["seg"] = None
-            logging.warning(f"Segmentation unavailable: {e}")
+            logging.warning(f"Segmentation model not available: {e}")
 
     if "cls" not in models_cache:
         try:
             models_cache["cls"] = load_classification_pipeline()
-            logging.info("✅ Classifier loaded (CPU)")
+            logging.info("✅ Classification pipeline loaded (CPU)")
         except Exception as e:
             models_cache["cls"] = None
-            logging.warning(f"Classifier unavailable: {e}")
+            logging.warning(f"Classification pipeline not available: {e}")
 
     if "embedding_model" not in models_cache:
         try:
             models_cache["embedding_model"] = load_embedding_model()
-            logging.info("✅ Embeddings loaded (CPU)")
+            logging.info("✅ Embedding model loaded (CPU)")
         except Exception as e:
             models_cache["embedding_model"] = None
-            logging.warning(f"Embeddings unavailable: {e}")
+            logging.warning(f"Embedding model not available: {e}")
 
 def setup_knowledge_base() -> None:
+    """Load PDFs and create FAISS vector store (optional)."""
     if "vector_store" in knowledge_base_cache:
         return
-    docs: List = []
+
+    docs = []
     try:
         PyPDFLoader = _import_langchain_pdf()
         for pdf in GUIDELINE_PDFS:
             if os.path.exists(pdf):
                 try:
-                    docs.extend(PyPDFLoader(pdf).load())
+                    loader = PyPDFLoader(pdf)
+                    docs.extend(loader.load())
                     logging.info(f"Loaded PDF: {pdf}")
                 except Exception as e:
-                    logging.warning(f"PDF load failed ({pdf}): {e}")
+                    logging.warning(f"Failed to load PDF {pdf}: {e}")
     except Exception as e:
         logging.warning(f"LangChain PDF loader unavailable: {e}")
 
@@ -218,603 +251,145 @@ def setup_knowledge_base() -> None:
         try:
             from langchain.text_splitter import RecursiveCharacterTextSplitter
             FAISS = _import_langchain_faiss()
-            chunks = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100).split_documents(docs)
+            splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
+            chunks = splitter.split_documents(docs)
             knowledge_base_cache["vector_store"] = FAISS.from_documents(chunks, models_cache["embedding_model"])
-            logging.info(f"✅ Knowledge base ready ({len(chunks)} chunks)")
+            logging.info(f"✅ Knowledge base ready with {len(chunks)} chunks")
         except Exception as e:
             knowledge_base_cache["vector_store"] = None
-            logging.warning(f"KB build failed: {e}")
+            logging.warning(f"Knowledge base unavailable: {e}")
     else:
         knowledge_base_cache["vector_store"] = None
-        logging.warning("KB disabled (no docs or embeddings).")
+        logging.warning("Knowledge base disabled (no docs or embeddings).")
 
+# Initialize on import
 initialize_cpu_models()
 setup_knowledge_base()
 
-# ---------- Calibration helpers ----------
-def _adaptive_prob_threshold(p: np.ndarray) -> float:
-    """
-    Pick a threshold that avoids tiny blobs while not swallowing skin.
-    Strategy:
-      - try Otsu on the prob map
-      - clamp to a reasonable band [0.25, 0.65]
-      - also consider percentile cut (p90) and take the "best" by area heuristic
-    """
-    p01 = np.clip(p.astype(np.float32), 0, 1)
-    p255 = (p01 * 255).astype(np.uint8)
-
-    # Otsu → use the returned scalar threshold (ret), NOT the image
-    ret_otsu, _dst = cv2.threshold(p255, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
-    thr_otsu = float(np.clip(ret_otsu / 255.0, 0.25, 0.65))
-
-    # Percentile (90th)
-    thr_pctl = float(np.clip(np.percentile(p01, 90), 0.25, 0.65))
-
-    # Area fraction helper
-    def area_frac(thr: float) -> float:
-        return float((p01 >= thr).sum()) / float(p01.size)
-
-    af_otsu = area_frac(thr_otsu)
-    af_pctl = area_frac(thr_pctl)
-
-    # Score: prefer ~3–10% coverage
-    def score(af: float) -> float:
-        target_low, target_high = 0.03, 0.10
-        if af < target_low: return abs(af - target_low) * 3.0
-        if af > target_high: return abs(af - target_high) * 1.5
-        return 0.0
-
-    return thr_otsu if score(af_otsu) <= score(af_pctl) else thr_pctl
-
-
-    # Score: closeness to a target area fraction (aim ~3–10%)
-    def score(af):
-        target_low, target_high = 0.03, 0.10
-        if af < target_low: return abs(af - target_low) * 3.0
-        if af > target_high: return abs(af - target_high) * 1.5
-        return 0.0
-
-    return thr_otsu if score(af_otsu) <= score(af_pctl) else thr_pctl
-
-
-def _grabcut_refine(bgr: np.ndarray, seed01: np.ndarray, iters: int = 3) -> np.ndarray:
-    """
-    Use OpenCV GrabCut to grow from a confident core into low-contrast margins.
-    seed01: 1=probable FG core, 0=unknown/other
-    """
-    h, w = bgr.shape[:2]
-    # Build GC mask: start with "unknown"
-    gc = np.full((h, w), cv2.GC_PR_BGD, np.uint8)
-    # definite FG = dilated seed; probable FG = seed
-    k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
-    seed_dil = cv2.dilate(seed01, k, iterations=1)
-    gc[seed01.astype(bool)] = cv2.GC_PR_FGD
-    gc[seed_dil.astype(bool)] = cv2.GC_FGD
-    # border is probable background
-    gc[0, :], gc[-1, :], gc[:, 0], gc[:, -1] = cv2.GC_BGD, cv2.GC_BGD, cv2.GC_BGD, cv2.GC_BGD
-
-    bgdModel = np.zeros((1, 65), np.float64)
-    fgdModel = np.zeros((1, 65), np.float64)
-    cv2.grabCut(bgr, gc, None, bgdModel, fgdModel, iters, cv2.GC_INIT_WITH_MASK)
-
-    # FG = definite or probable foreground
-    mask01 = np.where((gc == cv2.GC_FGD) | (gc == cv2.GC_PR_FGD), 1, 0).astype(np.uint8)
-    return mask01
-
-
-def _fill_holes(mask01: np.ndarray) -> np.ndarray:
-    h, w = mask01.shape[:2]
-    ff = np.zeros((h + 2, w + 2), np.uint8)
-    m = (mask01 * 255).astype(np.uint8).copy()
-    cv2.floodFill(m, ff, (0, 0), 255)
-    m_inv = cv2.bitwise_not(m)
-    out = ((mask01 * 255) | m_inv) // 255
-    return out.astype(np.uint8)
-
-
-def _clean_mask(mask01: np.ndarray) -> np.ndarray:
-    """Open → Close → Fill holes → Largest component → light smooth."""
-    mask01 = (mask01 > 0).astype(np.uint8)
-    k3 = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
-    k5 = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
-    mask01 = cv2.morphologyEx(mask01, cv2.MORPH_OPEN, k3, iterations=1)
-    mask01 = cv2.morphologyEx(mask01, cv2.MORPH_CLOSE, k5, iterations=2)
-    mask01 = _fill_holes(mask01)
-
-    # keep largest component
-    num, labels, stats, _ = cv2.connectedComponentsWithStats(mask01, 8)
-    if num > 1:
-        areas = stats[1:, cv2.CC_STAT_AREA]
-        if areas.size:
-            largest_idx = 1 + int(np.argmax(areas))
-            mask01 = (labels == largest_idx).astype(np.uint8)
-
-    # tiny masks → gentle grow (distance transform based)
-    area = int(mask01.sum())
-    if area > 0:
-        grow = 1 if area < 2000 else 0
-        if grow:
-            k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
-            mask01 = cv2.dilate(mask01, k, iterations=1)
-
-    return (mask01 > 0).astype(np.uint8)
-
-
-
-
-
-def _exif_to_dict(pil_img: Image.Image) -> Dict[str, object]:
-    out = {}
-    try:
-        exif = pil_img.getexif()
-        if not exif:
-            return out
-        for k, v in exif.items():
-            tag = TAGS.get(k, k)
-            out[tag] = v
-    except Exception:
-        pass
-    return out
-
-def _to_float(val) -> Optional[float]:
-    try:
-        if val is None:
-            return None
-        if isinstance(val, tuple) and len(val) == 2:
-            num, den = float(val[0]), float(val[1]) if float(val[1]) != 0 else 1.0
-            return num / den
-        return float(val)
-    except Exception:
-        return None
-
-def _estimate_sensor_width_mm(f_mm: Optional[float], f35: Optional[float]) -> Optional[float]:
-    if f_mm and f35 and f35 > 0:
-        return 36.0 * f_mm / f35
-    return None
-
-def estimate_px_per_cm_from_exif(pil_img: Image.Image, default_px_per_cm: float = DEFAULT_PX_PER_CM) -> Tuple[float, Dict]:
-    meta = {"used": "default", "f_mm": None, "f35": None, "sensor_w_mm": None, "distance_m": None}
-    try:
-        exif = _exif_to_dict(pil_img)
-        f_mm = _to_float(exif.get("FocalLength"))
-        f35 = _to_float(exif.get("FocalLengthIn35mmFilm") or exif.get("FocalLengthIn35mm"))
-        subj_dist_m = _to_float(exif.get("SubjectDistance"))
-        sensor_w_mm = _estimate_sensor_width_mm(f_mm, f35)
-        meta.update({"f_mm": f_mm, "f35": f35, "sensor_w_mm": sensor_w_mm, "distance_m": subj_dist_m})
-
-        if f_mm and sensor_w_mm and subj_dist_m and subj_dist_m > 0:
-            w_px = pil_img.width
-            field_w_mm = sensor_w_mm * (subj_dist_m * 1000.0) / f_mm
-            field_w_cm = field_w_mm / 10.0
-            px_per_cm = w_px / max(field_w_cm, 1e-6)
-            px_per_cm = float(np.clip(px_per_cm, PX_PER_CM_MIN, PX_PER_CM_MAX))
-            meta["used"] = "exif"
-            return px_per_cm, meta
-        return float(default_px_per_cm), meta
-    except Exception:
-        return float(default_px_per_cm), meta
-
-# ---------- Segmentation helpers ----------
-def _imagenet_norm(arr: np.ndarray) -> np.ndarray:
-    # expects RGB 0..255 -> float
-    mean = np.array([123.675, 116.28, 103.53], dtype=np.float32)
-    std  = np.array([58.395, 57.12, 57.375], dtype=np.float32)
-    return (arr.astype(np.float32) - mean) / std
-
-def _preprocess_for_seg(bgr_roi: np.ndarray, target_hw: Tuple[int, int]) -> np.ndarray:
-    H, W = target_hw
-    resized = cv2.resize(bgr_roi, (W, H), interpolation=cv2.INTER_LINEAR)
-    if SEG_EXPECTS_RGB:
-        resized = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB)
-    if SEG_NORM.lower() == "imagenet":
-        x = _imagenet_norm(resized)
-    else:
-        x = resized.astype(np.float32) / 255.0
-    x = np.expand_dims(x, axis=0)  # (1,H,W,3)
-    return x
-
-def _to_prob(pred: np.ndarray) -> np.ndarray:
-    p = np.squeeze(pred)
-    pmin, pmax = float(p.min()), float(p.max())
-    if pmax > 1.0 or pmin < 0.0:
-        p = 1.0 / (1.0 + np.exp(-p))
-    return p.astype(np.float32)
-
-# ---- Robust mask post-processing (for "proper" masking) ----
-def _fill_holes(mask01: np.ndarray) -> np.ndarray:
-    # Flood-fill from border, then invert
-    h, w = mask01.shape[:2]
-    ff = np.zeros((h + 2, w + 2), np.uint8)
-    m = (mask01 * 255).astype(np.uint8).copy()
-    cv2.floodFill(m, ff, (0, 0), 255)
-    m_inv = cv2.bitwise_not(m)
-    # Combine original with filled holes
-    out = ((mask01 * 255) | m_inv) // 255
-    return out.astype(np.uint8)
-
-# Global last debug dict (per-process) to attach into results
-_last_seg_debug: Dict[str, object] = {}
-
-def segment_wound(image_bgr: np.ndarray, ts: str, out_dir: str) -> Tuple[np.ndarray, Dict[str, object]]:
-    """
-    TF model → adaptive threshold on prob → (optional) GrabCut grow → cleanup.
-    Falls back to KMeans-Lab when model missing/fails.
-    Returns (mask_uint8_0_255, debug_dict)
-    """
-    debug = {"used": None, "reason": None, "positive_fraction": 0.0,
-             "thr": None, "heatmap_path": None, "roi_seen_by_model": None}
-
-    seg_model = models_cache.get("seg", None)
-
-    # --- Model path ---
-    if seg_model is not None:
-        try:
-            ishape = getattr(seg_model, "input_shape", None)
-            if not ishape or len(ishape) < 4:
-                raise ValueError(f"Bad seg input_shape: {ishape}")
-            th, tw = int(ishape[1]), int(ishape[2])
-
-            # preprocess
-            x = _preprocess_for_seg(image_bgr, (th, tw))
-            rgb_for_view = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
-            roi_seen_path = None
-            if SMARTHEAL_DEBUG:
-                roi_seen_path = os.path.join(out_dir, f"roi_for_seg_{ts}.png")
-                cv2.imwrite(roi_seen_path, cv2.cvtColor(rgb_for_view, cv2.COLOR_RGB2BGR))
-
-            # predict → prob map back to ROI size
-            pred = seg_model.predict(x, verbose=0)
-            if isinstance(pred, (list, tuple)): pred = pred[0]
-            p = _to_prob(pred)
-            p = cv2.resize(p, (image_bgr.shape[1], image_bgr.shape[0]), interpolation=cv2.INTER_LINEAR)
-
-            # visualization (optional)
-            heatmap_path = None
-            if SMARTHEAL_DEBUG:
-                hm = (np.clip(p, 0, 1) * 255).astype(np.uint8)
-                heat = cv2.applyColorMap(hm, cv2.COLORMAP_JET)
-                heatmap_path = os.path.join(out_dir, f"seg_pred_heatmap_{ts}.png")
-                cv2.imwrite(heatmap_path, heat)
-
-            # --- Adaptive threshold ---
-            thr = _adaptive_prob_threshold(p)
-            core01 = (p >= thr).astype(np.uint8)
-            core_frac = float(core01.sum()) / float(core01.size)
-
-            # If still too tiny, try a gentler threshold
-            if core_frac < 0.005:
-                thr2 = max(thr - 0.10, 0.15)
-                core01 = (p >= thr2).astype(np.uint8)
-                thr = thr2
-                core_frac = float(core01.sum()) / float(core01.size)
-
-            # --- Grow with GrabCut (only if some core exists) ---
-            if core01.any():
-                gc01 = _grabcut_refine(image_bgr, core01, iters=3)
-                mask01 = _clean_mask(gc01)
-            else:
-                mask01 = np.zeros(core01.shape, np.uint8)
-
-            pos_frac = float(mask01.sum()) / float(mask01.size)
-            logging.info(f"SegModel USED | thr={thr:.2f} core_frac={core_frac:.4f} final_frac={pos_frac:.4f}")
-
-            debug.update({
-                "used": "tf_model",
-                "reason": "ok",
-                "positive_fraction": pos_frac,
-                "thr": thr,
-                "heatmap_path": heatmap_path,
-                "roi_seen_by_model": roi_seen_path
-            })
-            return (mask01 * 255).astype(np.uint8), debug
-
-        except Exception as e:
-            logging.warning(f"⚠️ Segmentation model failed → fallback. Reason: {e}")
-            debug.update({"used": "fallback_kmeans", "reason": f"model_failed: {e}"})
-
-    # --- Fallback: KMeans in Lab (reddest cluster as wound) ---
-    Z = image_bgr.reshape((-1, 3)).astype(np.float32)
-    criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 10, 1.0)
-    _, labels, centers = cv2.kmeans(Z, 2, None, criteria, 5, cv2.KMEANS_PP_CENTERS)
-    centers_u8 = centers.astype(np.uint8).reshape(1, 2, 3)
-    centers_lab = cv2.cvtColor(centers_u8, cv2.COLOR_BGR2LAB)[0]
-    wound_idx = int(np.argmax(centers_lab[:, 1]))  # maximize a* (red)
-    mask01 = (labels.reshape(image_bgr.shape[:2]) == wound_idx).astype(np.uint8)
-    mask01 = _clean_mask(mask01)
-
-    pos_frac = float(mask01.sum()) / float(mask01.size)
-    logging.info(f"KMeans USED | final_frac={pos_frac:.4f}")
-
-    debug.update({
-        "used": "fallback_kmeans",
-        "reason": debug.get("reason") or "no_model",
-        "positive_fraction": pos_frac,
-        "thr": None
-    })
-    return (mask01 * 255).astype(np.uint8), debug
-
-
-# ---------- Measurement + overlay helpers ----------
-def largest_component_mask(binary01: np.ndarray, min_area_px: int = 50) -> np.ndarray:
-    num, labels, stats, _ = cv2.connectedComponentsWithStats(binary01.astype(np.uint8), connectivity=8)
-    if num <= 1:
-        return binary01.astype(np.uint8)
-    areas = stats[1:, cv2.CC_STAT_AREA]
-    if areas.size == 0 or areas.max() < min_area_px:
-        return binary01.astype(np.uint8)
-    largest_idx = 1 + int(np.argmax(areas))
-    return (labels == largest_idx).astype(np.uint8)
-
-def _clean_mask(mask01: np.ndarray) -> np.ndarray:
-    """Open→Close→Fill holes→Largest component."""
-    if mask01.dtype != np.uint8:
-        mask01 = mask01.astype(np.uint8)
-    k = np.ones((3, 3), np.uint8)
-    mask01 = cv2.morphologyEx(mask01, cv2.MORPH_OPEN, k, iterations=1)
-    mask01 = cv2.morphologyEx(mask01, cv2.MORPH_CLOSE, k, iterations=2)
-    mask01 = _fill_holes(mask01)
-    mask01 = largest_component_mask(mask01, min_area_px=30)
-    return (mask01 > 0).astype(np.uint8)
-
-def measure_min_area_rect(mask01: np.ndarray, px_per_cm: float) -> Tuple[float, float, Tuple]:
-    contours, _ = cv2.findContours(mask01.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-    if not contours:
-        return 0.0, 0.0, (None, None)
-    cnt = max(contours, key=cv2.contourArea)
-    rect = cv2.minAreaRect(cnt)
-    (w_px, h_px) = rect[1]
-    length_px, breadth_px = (max(w_px, h_px), min(w_px, h_px))
-    length_cm = round(length_px / max(px_per_cm, 1e-6), 2)
-    breadth_cm = round(breadth_px / max(px_per_cm, 1e-6), 2)
-    box = cv2.boxPoints(rect).astype(int)
-    return length_cm, breadth_cm, (box, rect[0])
-
-def count_area_cm2(mask01: np.ndarray, px_per_cm: float) -> float:
-    px_count = float(mask01.astype(bool).sum())
-    return round(px_count / (max(px_per_cm, 1e-6) ** 2), 2)
-
-def draw_measurement_overlay(
-    base_bgr: np.ndarray,
-    mask01: np.ndarray,
-    rect_box: np.ndarray,
-    length_cm: float,
-    breadth_cm: float,
-    thickness: int = 2
-) -> np.ndarray:
-    """
-    Draws:
-      1) Strong red mask overlay with white contour.
-      2) Min-area rectangle.
-      3) Two double-headed arrows:
-           - 'Length' along the longer side.
-           - 'Width'  along the shorter side.
-    """
-    overlay = base_bgr.copy()
-
-    # --- Strong overlay from mask (tinted red where mask==1) ---
-    mask255 = (mask01 * 255).astype(np.uint8)
-    mask3 = cv2.merge([mask255, mask255, mask255])
-    red = np.zeros_like(overlay); red[:] = (0, 0, 255)
-    alpha = 0.55
-    tinted = cv2.addWeighted(overlay, 1 - alpha, red, alpha, 0)
-    overlay = np.where(mask3 > 0, tinted, overlay)
-
-    # Draw wound contour
-    cnts, _ = cv2.findContours(mask255, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-    if cnts:
-        cv2.drawContours(overlay, cnts, -1, (255, 255, 255), 2)
-
-    if rect_box is not None:
-        cv2.polylines(overlay, [rect_box], True, (255, 255, 255), thickness)
-        pts = rect_box.reshape(-1, 2)
-
-        def midpoint(a, b):
-            return (int((a[0] + b[0]) / 2), int((a[1] + b[1]) / 2))
-
-        # Edge lengths
-        e = [np.linalg.norm(pts[i] - pts[(i + 1) % 4]) for i in range(4)]
-        long_edge_idx = int(np.argmax(e))
-        short_edge_idx = (long_edge_idx + 1) % 2  # 0/1 map for pairs below
-
-        # Midpoints of opposite edges for arrows
-        mids = [midpoint(pts[i], pts[(i + 1) % 4]) for i in range(4)]
-        # Long side uses edges long_edge_idx and the opposite edge (i+2)
-        long_pair = (long_edge_idx, (long_edge_idx + 2) % 4)
-        # Short side uses the other pair
-        short_pair = ((long_edge_idx + 1) % 4, (long_edge_idx + 3) % 4)
-
-        def draw_double_arrow(img, p1, p2):
-            cv2.arrowedLine(img, p1, p2, (0, 0, 0), thickness + 2, tipLength=0.05)
-            cv2.arrowedLine(img, p2, p1, (0, 0, 0), thickness + 2, tipLength=0.05)
-            cv2.arrowedLine(img, p1, p2, (255, 255, 255), thickness, tipLength=0.05)
-            cv2.arrowedLine(img, p2, p1, (255, 255, 255), thickness, tipLength=0.05)
-
-        def put_label(text, anchor):
-            org = (anchor[0] + 6, anchor[1] - 6)
-            cv2.putText(overlay, text, org, cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 0), 4, cv2.LINE_AA)
-            cv2.putText(overlay, text, org, cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 255), 2, cv2.LINE_AA)
-
-        # Draw arrows and labels
-        draw_double_arrow(overlay, mids[long_pair[0]], mids[long_pair[1]])
-        draw_double_arrow(overlay, mids[short_pair[0]], mids[short_pair[1]])
-        put_label(f"Length: {length_cm:.2f} cm", mids[long_pair[0]])
-        put_label(f"Width:  {breadth_cm:.2f} cm", mids[short_pair[0]])
-
-    return overlay
-
-# ---------- AI PROCESSOR ----------
+# =============== AI PROCESSOR ===============
 class AIProcessor:
     def __init__(self):
         self.models_cache = models_cache
         self.knowledge_base_cache = knowledge_base_cache
+        self.px_per_cm = PIXELS_PER_CM
         self.uploads_dir = UPLOADS_DIR
         self.dataset_id = DATASET_ID
         self.hf_token = HF_TOKEN
 
+    # ---------- Image utilities ----------
     def _ensure_analysis_dir(self) -> str:
         out_dir = os.path.join(self.uploads_dir, "analysis")
         os.makedirs(out_dir, exist_ok=True)
         return out_dir
 
     def perform_visual_analysis(self, image_pil: Image.Image) -> Dict:
-        """
-        YOLO detect → crop ROI → segment_wound(ROI) → clean mask →
-        minAreaRect measurement (cm) using EXIF px/cm → save outputs.
-        """
+        """YOLO detect → (optional) Keras seg → (optional) HF classifier → save visuals."""
         try:
-            px_per_cm, exif_meta = estimate_px_per_cm_from_exif(image_pil, DEFAULT_PX_PER_CM)
             image_cv = cv2.cvtColor(np.array(image_pil.convert("RGB")), cv2.COLOR_RGB2BGR)
 
-            # --- Detection ---
-            det_model = self.models_cache.get("det")
-            if det_model is None:
+            det = self.models_cache.get("det")
+            if det is None:
                 raise RuntimeError("YOLO model not loaded")
-            results = det_model.predict(image_cv, verbose=False, device="cpu")
-            if (not results) or (not getattr(results[0], "boxes", None)) or (len(results[0].boxes) == 0):
-                try:
-                    import gradio as gr
-                    raise gr.Error("No wound could be detected.")
-                except Exception:
-                    raise RuntimeError("No wound could be detected.")
+
+            # YOLO on CPU
+            results = det.predict(image_cv, verbose=False, device="cpu")
+            if not results or not getattr(results[0], "boxes", None) or len(results[0].boxes) == 0:
+                raise ValueError("No wound could be detected.")
 
             box = results[0].boxes[0].xyxy[0].cpu().numpy().astype(int)
             x1, y1, x2, y2 = [int(v) for v in box]
             x1, y1 = max(0, x1), max(0, y1)
             x2, y2 = min(image_cv.shape[1], x2), min(image_cv.shape[0], y2)
-            roi = image_cv[y1:y2, x1:x2].copy()
-            if roi.size == 0:
-                try:
-                    import gradio as gr
-                    raise gr.Error("Detected ROI is empty.")
-                except Exception:
-                    raise RuntimeError("Detected ROI is empty.")
-
-            out_dir = self._ensure_analysis_dir()
-            ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+            detected_region_cv = image_cv[y1:y2, x1:x2]
 
-            # --- Segmentation (model-first + KMeans fallback) ---
-            mask_u8_255, seg_debug = segment_wound(roi, ts, out_dir)
-            mask01 = (mask_u8_255 > 127).astype(np.uint8)
-
-            # Robust post-processing to ensure "proper" masking
-            if mask01.any():
-                mask01 = _clean_mask(mask01)
-                logging.debug(f"Mask postproc: px_after={int(mask01.sum())}")
-
-            # --- Measurement ---
-            if mask01.any():
-                length_cm, breadth_cm, (box_pts, _) = measure_min_area_rect(mask01, px_per_cm)
-                surface_area_cm2 = count_area_cm2(mask01, px_per_cm)
-                # Final annotated ROI with mask + arrows + labels
-                anno_roi = draw_measurement_overlay(roi, mask01, box_pts, length_cm, breadth_cm)
-                segmentation_empty = False
-            else:
-                # Graceful fallback if seg failed: use ROI box as bounds
-                h_px = max(0, y2 - y1); w_px = max(0, x2 - x1)
-                length_cm = round(max(h_px, w_px) / px_per_cm, 2)
-                breadth_cm = round(min(h_px, w_px) / px_per_cm, 2)
-                surface_area_cm2 = round((h_px * w_px) / (px_per_cm ** 2), 2)
-                anno_roi = roi.copy()
-                cv2.rectangle(anno_roi, (2, 2), (anno_roi.shape[1]-3, anno_roi.shape[0]-3), (0, 0, 255), 3)
-                cv2.line(anno_roi, (0, 0), (anno_roi.shape[1]-1, anno_roi.shape[0]-1), (0, 0, 255), 2)
-                cv2.line(anno_roi, (anno_roi.shape[1]-1, 0), (0, anno_roi.shape[0]-1), (0, 0, 255), 2)
-                box_pts = None
-                segmentation_empty = True
-
-            # --- Save visualizations ---
-            original_path = os.path.join(out_dir, f"original_{ts}.png")
-            cv2.imwrite(original_path, image_cv)
-
-            det_vis = image_cv.copy()
-            cv2.rectangle(det_vis, (x1, y1), (x2, y2), (0, 255, 0), 2)
-            detection_path = os.path.join(out_dir, f"detection_{ts}.png")
-            cv2.imwrite(detection_path, det_vis)
-
-            roi_mask_path = os.path.join(out_dir, f"roi_mask_{ts}.png")
-            cv2.imwrite(roi_mask_path, (mask01 * 255).astype(np.uint8))
-
-            # ROI overlay (clear mask w/ white contour, no arrows)
-            mask255 = (mask01 * 255).astype(np.uint8)
-            mask3   = cv2.merge([mask255, mask255, mask255])
-            red     = np.zeros_like(roi); red[:] = (0, 0, 255)
-            alpha   = 0.55
-            tinted  = cv2.addWeighted(roi, 1 - alpha, red, alpha, 0)
-            if mask255.any():
-                roi_overlay = np.where(mask3 > 0, tinted, roi)
-                cnts, _ = cv2.findContours(mask255, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-                cv2.drawContours(roi_overlay, cnts, -1, (255, 255, 255), 2)
-            else:
-                roi_overlay = anno_roi
-
-            seg_full = image_cv.copy()
-            seg_full[y1:y2, x1:x2] = roi_overlay
-            segmentation_path = os.path.join(out_dir, f"segmentation_{ts}.png")
-            cv2.imwrite(segmentation_path, seg_full)
-
-            segmentation_roi_path = os.path.join(out_dir, f"segmentation_roi_{ts}.png")
-            cv2.imwrite(segmentation_roi_path, roi_overlay)
-
-            # Annotated (mask + arrows + labels) in full-frame
-            anno_full = image_cv.copy()
-            anno_full[y1:y2, x1:x2] = anno_roi
-            annotated_seg_path = os.path.join(out_dir, f"segmentation_annotated_{ts}.png")
-            cv2.imwrite(annotated_seg_path, anno_full)
+            # Optional segmentation
+            seg_model = self.models_cache.get("seg")
+            length = breadth = area = 0.0
+            seg_path = None
+            if seg_model is not None and detected_region_cv.size > 0:
+                try:
+                    input_size = seg_model.input_shape[1:3]
+                    resized = cv2.resize(detected_region_cv, (input_size[1], input_size[0]))
+                    mask_pred = seg_model.predict(np.expand_dims(resized / 255.0, 0), verbose=0)[0]
+                    mask_np = (mask_pred[:, :, 0] > 0.5).astype(np.uint8)
+
+                    contours, _ = cv2.findContours(mask_np, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+                    if contours:
+                        cnt = max(contours, key=cv2.contourArea)
+                        x, y, w, h = cv2.boundingRect(cnt)
+                        length = round(h / self.px_per_cm, 2)
+                        breadth = round(w / self.px_per_cm, 2)
+                        area = round(cv2.contourArea(cnt) / (self.px_per_cm ** 2), 2)
+
+                        # overlay visualization
+                        mask_resized = cv2.resize(
+                            mask_np * 255,
+                            (detected_region_cv.shape[1], detected_region_cv.shape[0]),
+                            interpolation=cv2.INTER_NEAREST,
+                        )
+                        overlay = detected_region_cv.copy()
+                        overlay[mask_resized > 127] = [0, 0, 255]
+                        seg_vis = cv2.addWeighted(detected_region_cv, 0.7, overlay, 0.3, 0)
+
+                        ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+                        out_dir = self._ensure_analysis_dir()
+                        seg_path = os.path.join(out_dir, f"segmentation_{ts}.png")
+                        cv2.imwrite(seg_path, seg_vis)
+                except Exception as e:
+                    logging.warning(f"Segmentation step skipped: {e}")
 
-            # --- Optional classification ---
+            # Optional classification
             wound_type = "Unknown"
             cls_pipe = self.models_cache.get("cls")
             if cls_pipe is not None:
                 try:
-                    preds = cls_pipe(Image.fromarray(cv2.cvtColor(roi, cv2.COLOR_BGR2RGB)))
+                    detected_image_pil = Image.fromarray(cv2.cvtColor(detected_region_cv, cv2.COLOR_BGR2RGB))
+                    preds = cls_pipe(detected_image_pil)
                     if preds:
                         wound_type = max(preds, key=lambda x: x.get("score", 0)).get("label", "Unknown")
                 except Exception as e:
-                    logging.warning(f"Classification failed: {e}")
-
-            # Log end-of-seg summary
-            seg_summary = {
-                "seg_used": seg_debug.get("used"),
-                "seg_reason": seg_debug.get("reason"),
-                "positive_fraction": round(float(seg_debug.get("positive_fraction", 0.0)), 6),
-                "threshold": seg_debug.get("threshold", SEG_THRESH),
-                "segmentation_empty": segmentation_empty,
-                "exif_px_per_cm": round(px_per_cm, 3),
-            }
-            _log_kv("SEG_SUMMARY", seg_summary)
+                    logging.warning(f"Classification step failed: {e}")
+
+            # Save detection & original
+            out_dir = self._ensure_analysis_dir()
+            ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+            det_vis = image_cv.copy()
+            cv2.rectangle(det_vis, (x1, y1), (x2, y2), (0, 255, 0), 2)
+            det_path = os.path.join(out_dir, f"detection_{ts}.png")
+            cv2.imwrite(det_path, det_vis)
+
+            original_path = os.path.join(out_dir, f"original_{ts}.png")
+            cv2.imwrite(original_path, image_cv)
 
             return {
                 "wound_type": wound_type,
-                "length_cm": length_cm,
-                "breadth_cm": breadth_cm,
-                "surface_area_cm2": surface_area_cm2,
-                "px_per_cm": round(px_per_cm, 2),
-                "calibration_meta": exif_meta,
+                "length_cm": length,
+                "breadth_cm": breadth,
+                "surface_area_cm2": area,
                 "detection_confidence": float(results[0].boxes.conf[0].cpu().item())
-                    if getattr(results[0].boxes, "conf", None) is not None else 0.0,
-                "detection_image_path": detection_path,
-                "segmentation_image_path": annotated_seg_path,
-                "segmentation_annotated_path": annotated_seg_path,
-                "segmentation_roi_path": segmentation_roi_path,
-                "roi_mask_path": roi_mask_path,
-                "segmentation_empty": segmentation_empty,
-                "segmentation_debug": seg_debug,
+                if getattr(results[0].boxes, "conf", None) is not None
+                else 0.0,
+                "detection_image_path": det_path,
+                "segmentation_image_path": seg_path,
                 "original_image_path": original_path,
             }
         except Exception as e:
-            logging.error(f"Visual analysis failed: {e}", exc_info=True)
+            logging.error(f"Visual analysis failed: {e}")
             raise
 
-    # ---------- Knowledge base + reporting ----------
     def query_guidelines(self, query: str) -> str:
+        """Query the knowledge base (optional)."""
         try:
             vs = self.knowledge_base_cache.get("vector_store")
             if not vs:
                 return "Knowledge base is not available."
+            # support both old and new retriever APIs
             try:
                 retriever = vs.as_retriever(search_kwargs={"k": 5})
-                docs = retriever.get_relevant_documents(query)
+                docs = retriever.get_relevant_documents(query)  # LC >= 0.2
             except Exception:
                 retriever = vs.as_retriever(search_kwargs={"k": 5})
+                # older invoke API
                 docs = retriever.invoke(query)
             lines: List[str] = []
             for d in docs:
@@ -826,7 +401,9 @@ class AIProcessor:
             logging.warning(f"Guidelines query failed: {e}")
             return f"Guidelines query failed: {str(e)}"
 
+    # ---------- Report builders ----------
     def _generate_fallback_report(self, patient_info: str, visual_results: Dict, guideline_context: str) -> str:
+        """Plaintext/markdown fallback when MedGemma is unavailable."""
         return f"""# 🩺 SmartHeal AI - Comprehensive Wound Analysis Report
 ## 📋 Patient Information
 {patient_info}
@@ -835,12 +412,10 @@ class AIProcessor:
 - **Dimensions**: {visual_results.get('length_cm', 0)} cm × {visual_results.get('breadth_cm', 0)} cm
 - **Surface Area**: {visual_results.get('surface_area_cm2', 0)} cm²
 - **Detection Confidence**: {visual_results.get('detection_confidence', 0):.1%}
-- **Calibration**: {visual_results.get('px_per_cm','?')} px/cm ({(visual_results.get('calibration_meta') or {}).get('used','default')})
 ## 📊 Analysis Images
 - **Original**: {visual_results.get('original_image_path', 'N/A')}
 - **Detection**: {visual_results.get('detection_image_path', 'N/A')}
 - **Segmentation**: {visual_results.get('segmentation_image_path', 'N/A')}
-- **Annotated**: {visual_results.get('segmentation_annotated_path', 'N/A')}
 ## 🎯 Clinical Summary
 Automated analysis provides quantitative measurements; verify via clinical examination.
 ## 💊 Recommendations
@@ -848,10 +423,10 @@ Automated analysis provides quantitative measurements; verify via clinical exami
 - Debride necrotic tissue if indicated (clinical decision)
 - Document with serial photos and measurements
 ## 📅 Monitoring
-- Daily in week 1, then every 2–3 days (or as indicated)
+- Daily in week 1, then every 2-3 days (or as indicated)
 - Weekly progress review
 ## 📚 Guideline Context
-{(guideline_context or '')[:800]}{"..." if guideline_context and len(guideline_context) > 800 else ''}
+{(guideline_context or '')[:800]}{'...' if guideline_context and len(guideline_context) > 800 else ''}
 **Disclaimer:** Automated, for decision support only. Verify clinically.
 """
 
@@ -863,8 +438,9 @@ Automated analysis provides quantitative measurements; verify via clinical exami
         image_pil: Image.Image,
         max_new_tokens: Optional[int] = None,
     ) -> str:
+        """Try MedGemma (GPU) → fallback report."""
         try:
-            report = generate_medgemma_report(
+            report = generate_medgemma_report_with_timeout(
                 patient_info, visual_results, guideline_context, image_pil, max_new_tokens
             )
             if report and report.strip() and not report.startswith(("⚠️", "❌")):
@@ -875,7 +451,9 @@ Automated analysis provides quantitative measurements; verify via clinical exami
             logging.error(f"Report generation failed: {e}")
             return self._generate_fallback_report(patient_info, visual_results, guideline_context)
 
+    # ---------- HF dataset commit ----------
     def save_and_commit_image(self, image_pil: Image.Image) -> str:
+        """Save image locally and optionally upload to HF dataset."""
         try:
             os.makedirs(self.uploads_dir, exist_ok=True)
             ts = datetime.now().strftime("%Y%m%d_%H%M%S")
@@ -884,17 +462,17 @@ Automated analysis provides quantitative measurements; verify via clinical exami
             image_pil.convert("RGB").save(path)
             logging.info(f"✅ Image saved locally: {path}")
 
-            if HF_TOKEN and DATASET_ID:
+            if self.hf_token and self.dataset_id:
                 try:
                     HfApi, HfFolder = _import_hf_hub()
-                    HfFolder.save_token(HF_TOKEN)
+                    HfFolder.save_token(self.hf_token)
                     api = HfApi()
                     api.upload_file(
                         path_or_fileobj=path,
                         path_in_repo=f"images/{filename}",
-                        repo_id=DATASET_ID,
+                        repo_id=self.dataset_id,
                         repo_type="dataset",
-                        token=HF_TOKEN,
+                        token=self.hf_token,
                         commit_message=f"Upload wound image: {filename}",
                     )
                     logging.info("✅ Image committed to HF dataset")
@@ -906,23 +484,28 @@ Automated analysis provides quantitative measurements; verify via clinical exami
             logging.error(f"Failed to save/commit image: {e}")
             return ""
 
+    # ---------- Orchestrator ----------
     def full_analysis_pipeline(self, image_pil: Image.Image, questionnaire_data: Dict) -> Dict:
+        """End-to-end analysis with robust fallbacks."""
         try:
             saved_path = self.save_and_commit_image(image_pil)
+
             visual_results = self.perform_visual_analysis(image_pil)
 
+            # Patient info summary text
             pi = questionnaire_data or {}
             patient_info = (
-                f"Age: {pi.get('age','N/A')}, "
-                f"Diabetic: {pi.get('diabetic','N/A')}, "
-                f"Allergies: {pi.get('allergies','N/A')}, "
-                f"Date of Wound: {pi.get('date_of_injury','N/A')}, "
-                f"Professional Care: {pi.get('professional_care','N/A')}, "
-                f"Oozing/Bleeding: {pi.get('oozing_bleeding','N/A')}, "
-                f"Infection: {pi.get('infection','N/A')}, "
-                f"Moisture: {pi.get('moisture','N/A')}"
+                f"Age: {pi.get('age', 'N/A')}, "
+                f"Diabetic: {pi.get('diabetic', 'N/A')}, "
+                f"Allergies: {pi.get('allergies', 'N/A')}, "
+                f"Date of Wound: {pi.get('date_of_injury', 'N/A')}, "
+                f"Professional Care: {pi.get('professional_care', 'N/A')}, "
+                f"Oozing/Bleeding: {pi.get('oozing_bleeding', 'N/A')}, "
+                f"Infection: {pi.get('infection', 'N/A')}, "
+                f"Moisture: {pi.get('moisture', 'N/A')}"
             )
 
+            # Query guidelines
             query = (
                 f"best practices for managing a {visual_results.get('wound_type','Unknown')} "
                 f"with moisture '{pi.get('moisture','unknown')}' and infection '{pi.get('infection','unknown')}' "
@@ -930,16 +513,18 @@ Automated analysis provides quantitative measurements; verify via clinical exami
             )
             guideline_context = self.query_guidelines(query)
 
-            report = self.generate_final_report(patient_info, visual_results, guideline_context, image_pil)
+            # Generate final report
+            report = self.generate_final_report(patient_info=patient_info,
+                                                visual_results=visual_results,
+                                                guideline_context=guideline_context,
+                                                image_pil=image_pil)
 
             return {
                 "success": True,
                 "visual_analysis": visual_results,
                 "report": report,
                 "saved_image_path": saved_path,
-                "guideline_context": (guideline_context or "")[:500] + (
-                    "..." if guideline_context and len(guideline_context) > 500 else ""
-                ),
+                "guideline_context": (guideline_context or "")[:500] + ("..." if guideline_context and len(guideline_context) > 500 else ""),
             }
         except Exception as e:
             logging.error(f"Pipeline error: {e}")
@@ -953,6 +538,7 @@ Automated analysis provides quantitative measurements; verify via clinical exami
             }
 
     def analyze_wound(self, image, questionnaire_data: Dict) -> Dict:
+        """Public entrypoint used by your UI."""
         try:
             if isinstance(image, str):
                 if not os.path.exists(image):
@@ -975,4 +561,4 @@ Automated analysis provides quantitative measurements; verify via clinical exami
                 "report": f"Analysis initialization failed: {str(e)}",
                 "saved_image_path": None,
                 "guideline_context": "",
-            }
+            }