Update src/ai_processor.py

src/ai_processor.py

@@ -1,37 +1,53 @@
-# smartheal_ai_processor.py
-# Verbose, instrumented version — preserves public class/function names
-# Turn on deep logging: export LOGLEVEL=DEBUG SMARTHEAL_DEBUG=1
+"""
+SmartHeal AI Processor - Zero GPU Compatible Version
+Designed specifically for Hugging Face Spaces with Zero GPU architecture
+"""
 
 import os
 import logging
-from datetime import datetime
-from typing import Optional, Dict, List, Tuple
-
-# ---- Environment defaults (do NOT globally hint CUDA here) ----
-os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
-LOGLEVEL = os.getenv("LOGLEVEL", "INFO").upper()
-SMARTHEAL_DEBUG = os.getenv("SMARTHEAL_DEBUG", "0") == "1"
-
 import cv2
 import numpy as np
 from PIL import Image
-from PIL.ExifTags import TAGS
+import json
+from datetime import datetime
+from typing import Optional, Dict, List, Tuple, Any
+from contextlib import contextmanager
 
-# --- Logging config ---
+# Configure logging
 logging.basicConfig(
-    level=getattr(logging, LOGLEVEL, logging.INFO),
-    format="%(asctime)s - %(levelname)s - %(message)s",
+    level=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()))
-
-# --- Spaces GPU decorator (REQUIRED) ---
-from spaces import GPU as _SPACES_GPU
+# Environment setup for Zero GPU compatibility
+os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
+os.environ.setdefault("CUDA_VISIBLE_DEVICES", "-1")  # Hide GPU from main process
+
+# Import spaces decorator
+try:
+    import spaces
+    _SPACES_GPU = spaces.GPU
+except ImportError:
+    logging.warning("spaces package not available - running in CPU mode")
+    # Create dummy decorator for local testing
+    def _SPACES_GPU_dummy(*args, **kwargs):
+        def decorator(func):
+            return func
+        return decorator
+    _SPACES_GPU = _SPACES_GPU_dummy
 
-@_SPACES_GPU(enable_queue=True)
-def smartheal_gpu_stub(ping: int = 0) -> str:
-    return "ready"
+@contextmanager
+def _no_cuda_env():
+    """Context manager to prevent CUDA initialization in main process"""
+    prev_cuda = os.environ.get("CUDA_VISIBLE_DEVICES")
+    os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
+    try:
+        yield
+    finally:
+        if prev_cuda is None:
+            os.environ.pop("CUDA_VISIBLE_DEVICES", None)
+        else:
+            os.environ["CUDA_VISIBLE_DEVICES"] = prev_cuda
 
 # ---- Paths / constants ----
 UPLOADS_DIR = "uploads"
@@ -45,33 +61,9 @@ 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"))
-
 models_cache: Dict[str, object] = {}
 knowledge_base_cache: Dict[str, object] = {}
 
-# ---------- Utilities to prevent CUDA in main process ----------
-from contextlib import contextmanager
-
-@contextmanager
-def _no_cuda_env():
-    """
-    Mask GPUs so any library imported/constructed in the main process
-    cannot see CUDA (required for Spaces Stateless GPU).
-    """
-    prev = os.environ.get("CUDA_VISIBLE_DEVICES")
-    os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
-    try:
-        yield
-    finally:
-        if prev is None:
-            os.environ.pop("CUDA_VISIBLE_DEVICES", None)
-        else:
-            os.environ["CUDA_VISIBLE_DEVICES"] = prev
-
 # ---------- Lazy imports (wrapped where needed) ----------
 def _import_ultralytics():
     # Prevent Ultralytics from probing CUDA on import
@@ -80,33 +72,40 @@ def _import_ultralytics():
     return YOLO
 
 def _import_tf_loader():
-    import tensorflow as tf
-    tf.config.set_visible_devices([], "GPU")
-    from tensorflow.keras.models import load_model
+    # Ensure TensorFlow does not try to use GPU in main process
+    with _no_cuda_env():
+        import tensorflow as tf
+        tf.config.set_visible_devices([], "GPU")
+        from tensorflow.keras.models import load_model
     return load_model
 
 def _import_hf_cls():
-    from transformers import pipeline
+    with _no_cuda_env():
+        from transformers import pipeline
     return pipeline
 
 def _import_embeddings():
-    from langchain_community.embeddings import HuggingFaceEmbeddings
+    with _no_cuda_env():
+        from langchain_community.embeddings import HuggingFaceEmbeddings
     return HuggingFaceEmbeddings
 
 def _import_langchain_pdf():
-    from langchain_community.document_loaders import PyPDFLoader
+    with _no_cuda_env():
+        from langchain_community.document_loaders import PyPDFLoader
     return PyPDFLoader
 
 def _import_langchain_faiss():
-    from langchain_community.vectorstores import FAISS
+    with _no_cuda_env():
+        from langchain_community.vectorstores import FAISS
     return FAISS
 
 def _import_hf_hub():
-    from huggingface_hub import HfApi, HfFolder
+    with _no_cuda_env():
+        from huggingface_hub import HfApi, HfFolder
     return HfApi, HfFolder
 
 # ---------- SmartHeal prompts (system + user prefix) ----------
-SMARTHEAL_SYSTEM_PROMPT = """\
+SMARTHEAL_SYSTEM_PROMPT = """
 You are SmartHeal Clinical Assistant, a wound-care decision-support system.
 You analyze wound photographs and brief patient context to produce careful,
 specific, guideline-informed recommendations WITHOUT diagnosing. You always:
@@ -119,7 +118,7 @@ specific, guideline-informed recommendations WITHOUT diagnosing. You always:
 - Safety: remind the user to seek clinician review for changes or red flags.
 """
 
-SMARTHEAL_USER_PREFIX = """\
+SMARTHEAL_USER_PREFIX = """
 Patient: {patient_info}
 Visual findings: type={wound_type}, size={length_cm}x{breadth_cm} cm, area={area_cm2} cm^2,
 detection_conf={det_conf:.2f}, calibration={px_per_cm} px/cm.
@@ -139,6 +138,7 @@ Keep to 220–300 words. Do NOT provide diagnosis. Avoid contraindicated advice.
 """
 
 # ---------- VLM (MedGemma replaced with Qwen2-VL) ----------
+@_SPACES_GPU(enable_queue=True)
 def _vlm_infer_gpu(messages, model_id: str, max_new_tokens: int, token: Optional[str]):
     """
     Runs entirely inside a Spaces GPU worker. It's the ONLY place we allow CUDA init.
@@ -149,7 +149,7 @@ def _vlm_infer_gpu(messages, model_id: str, max_new_tokens: int, token: Optional
         task="image-text-to-text",
         model=model_id,
         torch_dtype=torch.bfloat16, # Use torch_dtype from the working example
-        device_map=0,            # CUDA init happens here, safely in GPU worker
+        device_map="auto",            # CUDA init happens here, safely in GPU worker
         token=token,
         trust_remote_code=True,
         model_kwargs={"low_cpu_mem_usage": True},
@@ -161,7 +161,6 @@ def _vlm_infer_gpu(messages, model_id: str, max_new_tokens: int, token: Optional
         txt = out[0].get("generated_text", "")
     return (txt or "").strip() or "⚠️ Empty response"
 
-@_SPACES_GPU(enable_queue=True)
 def generate_medgemma_report(  # kept name so callers don't change
     patient_info: str,
     visual_results: Dict,
@@ -213,9 +212,8 @@ def load_yolo_model():
         model = YOLO(YOLO_MODEL_PATH)
     return model
 def load_segmentation_model():
-    import tensorflow as tf
     load_model = _import_tf_loader()
-    return load_model(SEG_MODEL_PATH, compile=False, custom_objects={'InputLayer': tf.keras.layers.InputLayer})
+    return load_model(SEG_MODEL_PATH, compile=False, custom_objects={'InputLayer': _import_tf_loader().layers.InputLayer})
 
 def load_classification_pipeline():
     pipe = _import_hf_cls()
@@ -366,12 +364,7 @@ def _imagenet_norm(arr: np.ndarray) -> np.ndarray:
 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 = resized.astype(np.float32) / 255.0
     x = np.expand_dims(x, axis=0)  # (1,H,W,3)
     return x
 
@@ -410,7 +403,9 @@ def _adaptive_prob_threshold(p: np.ndarray) -> float:
     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:
-    """Grow from a confident core into low-contrast margins."""
+    """
+    Grow from a confident core into low-contrast margins.
+    """
     h, w = bgr.shape[:2]
     gc = np.full((h, w), cv2.GC_PR_BGD, np.uint8)
     k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
@@ -423,571 +418,127 @@ def _grabcut_refine(bgr: np.ndarray, seed01: np.ndarray, iters: int = 3) -> np.n
     cv2.grabCut(bgr, gc, None, bgdModel, fgdModel, iters, cv2.GC_INIT_WITH_MASK)
     return np.where((gc == cv2.GC_FGD) | (gc == cv2.GC_PR_FGD), 1, 0).astype(np.uint8)
 
-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 (no dilation)."""
-    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=1)
-    mask01 = _fill_holes(mask01)
-    # Keep largest component only
-    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)
-    return (mask01 > 0).astype(np.uint8)
-
-# Global last debug dict (per-process)
-_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 → GrabCut grow → cleanup.
-    Fallback: KMeans-Lab.
-    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])
-
-            x = _preprocess_for_seg(image_bgr, (th, tw))
-            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, image_bgr)
-
-            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)
-
-            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)
-
-            thr = _adaptive_prob_threshold(p)
-            core01 = (p >= thr).astype(np.uint8)
-            core_frac = float(core01.sum()) / float(core01.size)
-
-            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)
-
-            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={float(thr):.2f} core_frac={core_frac:.4f} final_frac={pos_frac:.4f}")
-
-            debug.update({
-                "used": "tf_model",
-                "reason": "ok",
-                "positive_fraction": pos_frac,
-                "thr": float(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 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 area_cm2_from_contour(mask01: np.ndarray, px_per_cm: float) -> Tuple[float, Optional[np.ndarray]]:
-    """Area from largest polygon (sub-pixel); returns (area_cm2, contour)."""
-    m = (mask01 > 0).astype(np.uint8)
-    contours, _ = cv2.findContours(m, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-    if not contours:
-        return 0.0, None
-    cnt = max(contours, key=cv2.contourArea)
-    poly_area_px2 = float(cv2.contourArea(cnt))
-    area_cm2 = round(poly_area_px2 / (max(px_per_cm, 1e-6) ** 2), 2)
-    return area_cm2, cnt
-
-def clamp_area_with_minrect(cnt: np.ndarray, px_per_cm: float, area_cm2_poly: float) -> float:
-    rect = cv2.minAreaRect(cnt)
-    (w_px, h_px) = rect[1]
-    rect_area_px2 = float(max(w_px, 0.0) * max(h_px, 0.0))
-    rect_area_cm2 = rect_area_px2 / (max(px_per_cm, 1e-6) ** 2)
-    return round(min(area_cm2_poly, rect_area_cm2 * 1.05), 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:
-    """
-    1) Strong red mask overlay + white contour
-    2) Min-area rectangle
-    3) Double-headed arrows labeled Length/Width
-    """
-    overlay = base_bgr.copy()
-
-    # Mask tint
-    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)
-
-    # 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))
-        e = [np.linalg.norm(pts[i] - pts[(i + 1) % 4]) for i in range(4)]
-        long_edge_idx = int(np.argmax(e))
-        mids = [midpoint(pts[i], pts[(i + 1) % 4]) for i in range(4)]
-        long_pair = (long_edge_idx, (long_edge_idx + 2) % 4)
-        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_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 ----------
+# ---------- Main AIProcessor Class ----------
 class AIProcessor:
     def __init__(self):
+        self.config = type("Config", (object,), {
+            "HF_TOKEN": HF_TOKEN,
+            "YOLO_MODEL_PATH": YOLO_MODEL_PATH,
+            "SEG_MODEL_PATH": SEG_MODEL_PATH,
+            "DATASET_ID": DATASET_ID,
+            "UPLOADS_DIR": UPLOADS_DIR,
+            "GUIDELINE_PDFS": GUIDELINE_PDFS
+        })()
         self.models_cache = models_cache
         self.knowledge_base_cache = knowledge_base_cache
-        self.uploads_dir = UPLOADS_DIR
-        self.dataset_id = DATASET_ID
-        self.hf_token = HF_TOKEN
-
-    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.
-        """
-        try:
-            px_per_cm, exif_meta = estimate_px_per_cm_from_exif(image_pil, DEFAULT_PX_PER_CM)
-            # Guardrails for calibration to avoid huge area blow-ups
-            px_per_cm = float(np.clip(px_per_cm, 20.0, 350.0))
-            if (exif_meta or {}).get("used") != "exif":
-                logging.warning(f"Calibration fallback used: px_per_cm={px_per_cm:.2f} (default). Prefer ruler/Aruco for accuracy.")
-
-            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:
-                raise RuntimeError("YOLO model not loaded")
-            # Force CPU inference and avoid CUDA touch
-            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.")
-
-            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")
-
-            # --- Segmentation (model-first + KMeans fallback) ---
-            mask_u8_255, seg_debug = segment_wound(roi, ts, out_dir)
-            mask01 = (mask_u8_255 > 127).astype(np.uint8)
-
-            if mask01.any():
-                mask01 = _clean_mask(mask01)
-                logging.debug(f"Mask postproc: px_after={int(mask01.sum())}")
-
-            # --- Measurement (accurate & conservative) ---
-            if mask01.any():
-                length_cm, breadth_cm, (box_pts, _) = measure_min_area_rect(mask01, px_per_cm)
-                area_poly_cm2, largest_cnt = area_cm2_from_contour(mask01, px_per_cm)
-                if largest_cnt is not None:
-                    surface_area_cm2 = clamp_area_with_minrect(largest_cnt, px_per_cm, area_poly_cm2)
-                else:
-                    surface_area_cm2 = area_poly_cm2
-
-                anno_roi = draw_measurement_overlay(roi, mask01, box_pts, length_cm, breadth_cm)
-                segmentation_empty = False
-            else:
-                # Fallback if seg failed: use ROI dimensions
-                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 (mask tint + contour, without 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 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)))
-                    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("thr"),
-                "segmentation_empty": segmentation_empty,
-                "exif_px_per_cm": round(px_per_cm, 3),
-            }
-            _log_kv("SEG_SUMMARY", seg_summary)
-
-            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,
-                "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,
-                "original_image_path": original_path,
-            }
-        except Exception as e:
-            logging.error(f"Visual analysis failed: {e}", exc_info=True)
-            raise
+        self.px_per_cm = DEFAULT_PX_PER_CM # Use default from constants
+
+        # Ensure CPU models and KB are initialized
+        initialize_cpu_models()
+        setup_knowledge_base()
+
+    def perform_visual_analysis(self, image_pil: Image.Image) -> Dict[str, Any]:
+        image_cv = cv2.cvtColor(np.array(image_pil), cv2.COLOR_RGB2BGR)
+
+        if "det" not in self.models_cache or not self.models_cache["det"]:
+            raise ValueError("YOLO model not initialized.")
+
+        results = self.models_cache["det"].predict(image_cv, verbose=False, device="cpu")
+        if not results or not results[0].boxes:
+            raise ValueError("No wound detected.")
+
+        box = results[0].boxes[0].xyxy[0].cpu().numpy().astype(int)
+        region_cv = image_cv[box[1]:box[3], box[0]:box[2]]
+        detection_confidence = float(results[0].boxes[0].conf[0].cpu().numpy())
+
+        length = breadth = area = 0
+        if "seg" in self.models_cache and self.models_cache["seg"]:
+            try:
+                seg_model = self.models_cache["seg"]
+                input_size = seg_model.input_shape[1:3]
+                preprocessed_roi = _preprocess_for_seg(region_cv, input_size)
+                mask_pred = seg_model.predict(preprocessed_roi, verbose=0)[0]
+                prob_mask = _to_prob(mask_pred)
+
+                # Adaptive thresholding and GrabCut refinement
+                initial_mask = (prob_mask >= _adaptive_prob_threshold(prob_mask)).astype(np.uint8)
+                refined_mask = _grabcut_refine(region_cv, initial_mask)
+
+                contours, _ = cv2.findContours(refined_mask, 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)
+            except Exception as e:
+                logging.warning(f"Segmentation process failed: {e}")
+
+        wound_type = "Unknown"
+        if "cls" in self.models_cache and self.models_cache["cls"]:
+            try:
+                wound_region_pil = Image.fromarray(cv2.cvtColor(region_cv, cv2.COLOR_BGR2RGB))
+                classification_results = self.models_cache["cls"](wound_region_pil)
+                wound_type = max(classification_results, key=lambda x: x["score"])["label"]
+            except Exception as e:
+                logging.warning(f"Classification process failed: {e}")
+
+        return {
+            "wound_type": wound_type,
+            "length_cm": length,
+            "breadth_cm": breadth,
+            "surface_area_cm2": area,
+            "detection_confidence": detection_confidence,
+            "px_per_cm": self.px_per_cm
+        }
 
-    # ---------- Knowledge base + reporting ----------
     def query_guidelines(self, query: str) -> str:
+        vector_store = self.knowledge_base_cache.get("vector_store")
+        if not vector_store:
+            return "Knowledge base unavailable."
+
+        retriever = vector_store.as_retriever(search_kwargs={"k": 10})
+        docs = retriever.invoke(query)
+        return "\n\n".join([
+            f"Source: {doc.metadata.get('source', 'N/A')}, Page: {doc.metadata.get('page', 'N/A')}\nContent: {doc.page_content}"
+            for doc in docs
+        ])
+
+    def generate_final_report(self, patient_info, visual_results, guideline_context, image_pil, max_new_tokens=2048):
+        return generate_medgemma_report(patient_info, visual_results, guideline_context, image_pil, max_new_tokens)
+
+    def save_and_commit_image(self, image_pil):
+        filename = f"{datetime.now().strftime('%Y-%m-%d_%H-%M-%S')}.png"
+        local_path = os.path.join(self.config.UPLOADS_DIR, filename)
+        image_pil.convert("RGB").save(local_path)
+        logging.info(f"Image saved locally: {local_path}")
+
+        if self.config.HF_TOKEN and self.config.DATASET_ID:
+            try:
+                api = _import_hf_hub()[0]() # HfApi
+                api.upload_file(
+                    path_or_fileobj=local_path,
+                    path_in_repo=f"images/{filename}",
+                    repo_id=self.config.DATASET_ID,
+                    repo_type="dataset",
+                    commit_message=f"Upload wound image: {filename}",
+                    token=self.config.HF_TOKEN
+                )
+                logging.info("✅ Image uploaded to HF dataset.")
+            except Exception as e:
+                logging.warning(f"Upload failed: {e}")
+
+    @_SPACES_GPU(enable_queue=True, duration=120)
+    def full_analysis_pipeline(self, image, questionnaire_data):
         try:
-            vs = self.knowledge_base_cache.get("vector_store")
-            if not vs:
-                return "Knowledge base is not available."
-            retriever = vs.as_retriever(search_kwargs={"k": 5})
-            # Modern API (avoid get_relevant_documents deprecation)
-            docs = retriever.invoke(query)
-            lines: List[str] = []
-            for d in docs:
-                src = (d.metadata or {}).get("source", "N/A")
-                txt = (d.page_content or "")[:300]
-                lines.append(f"Source: {src}\nContent: {txt}...")
-            return "\n\n".join(lines) if lines else "No relevant guideline snippets found."
-        except Exception as e:
-            logging.warning(f"Guidelines query failed: {e}")
-            return f"Guidelines query failed: {str(e)}"
-
-    def _generate_fallback_report(self, patient_info: str, visual_results: Dict, guideline_context: str) -> str:
-        return f"""# 🩺 SmartHeal AI - Comprehensive Wound Analysis Report
-
-## 📋 Patient Information
-{patient_info}
-
-## 🔍 Visual Analysis Results
-- **Wound Type**: {visual_results.get('wound_type', 'Unknown')}
-- **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
-- Cleanse wound gently; select dressing per exudate/infection risk
-- 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)
-- Weekly progress review
-
-## 📚 Guideline Context
-{(guideline_context or '')[:800]}{"..." if guideline_context and len(guideline_context) > 800 else ''}
-
-**Disclaimer:** Automated, for decision support only. Verify clinically.
-"""
-
-    def generate_final_report(
-        self,
-        patient_info: str,
-        visual_results: Dict,
-        guideline_context: str,
-        image_pil: Image.Image,
-        max_new_tokens: Optional[int] = None,
-    ) -> str:
-        try:
-            report = generate_medgemma_report(
-                patient_info, visual_results, guideline_context, image_pil, max_new_tokens
-            )
-            if report and report.strip() and not report.startswith(("⚠️", "❌")):
-                return report
-            logging.warning("VLM unavailable/invalid; using fallback.")
-            return self._generate_fallback_report(patient_info, visual_results, guideline_context)
-        except Exception as e:
-            logging.error(f"Report generation failed: {e}")
-            return self._generate_fallback_report(patient_info, visual_results, guideline_context)
-
-    def save_and_commit_image(self, image_pil: Image.Image) -> str:
-        try:
-            os.makedirs(self.uploads_dir, exist_ok=True)
-            ts = datetime.now().strftime("%Y%m%d_%H%M%S")
-            filename = f"{ts}.png"
-            path = os.path.join(self.uploads_dir, filename)
-            image_pil.convert("RGB").save(path)
-            logging.info(f"✅ Image saved locally: {path}")
-
-            if HF_TOKEN and DATASET_ID:
-                try:
-                    HfApi, HfFolder = _import_hf_hub()
-                    HfFolder.save_token(HF_TOKEN)
-                    api = HfApi()
-                    api.upload_file(
-                        path_or_fileobj=path,
-                        path_in_repo=f"images/{filename}",
-                        repo_id=DATASET_ID,
-                        repo_type="dataset",
-                        token=HF_TOKEN,
-                        commit_message=f"Upload wound image: {filename}",
-                    )
-                    logging.info("✅ Image committed to HF dataset")
-                except Exception as e:
-                    logging.warning(f"HF upload failed: {e}")
-
-            return path
-        except Exception as e:
-            logging.error(f"Failed to save/commit image: {e}")
-            return ""
-            
-    
-    def full_analysis_pipeline(self, image_pil: Image.Image, questionnaire_data: Dict) -> Dict:
-        try:
-            saved_path = self.save_and_commit_image(image_pil)
-            visual_results = self.perform_visual_analysis(image_pil)
-
-            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')}"
-            )
-
-            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')}' "
-                f"in a diabetic status '{pi.get('diabetic','unknown')}'"
-            )
+            self.save_and_commit_image(image)
+            visual = self.perform_visual_analysis(image)
+            patient_info = ", ".join([f"{k}: {v}" for k, v in questionnaire_data.items()])
+            query = f"best practices for managing a {visual['wound_type']} with moisture level '{questionnaire_data.get('moisture')}' and signs of infection '{questionnaire_data.get('infection')}' in a patient who is diabetic '{questionnaire_data.get('diabetic')}'"
             guideline_context = self.query_guidelines(query)
 
-            report = self.generate_final_report(patient_info, visual_results, guideline_context, 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 ""
-                ),
-            }
-        except Exception as e:
-            logging.error(f"Pipeline error: {e}")
-            return {
-                "success": False,
-                "error": str(e),
-                "visual_analysis": {},
-                "report": f"Analysis failed: {str(e)}",
-                "saved_image_path": None,
-                "guideline_context": "",
-            }
-
-    def analyze_wound(self, image, questionnaire_data: Dict) -> Dict:
-        try:
-            if isinstance(image, str):
-                if not os.path.exists(image):
-                    raise ValueError(f"Image file not found: {image}")
-                image_pil = Image.open(image)
-            elif isinstance(image, Image.Image):
-                image_pil = image
-            elif isinstance(image, np.ndarray):
-                image_pil = Image.fromarray(image)
-            else:
-                raise ValueError(f"Unsupported image type: {type(image)}")
+            return self.generate_final_report(patient_info, visual, guideline_context, image)
 
-            return self.full_analysis_pipeline(image_pil, questionnaire_data or {})
         except Exception as e:
-            logging.error(f"Wound analysis error: {e}")
-            return {
-                "success": False,
-                "error": str(e),
-                "visual_analysis": {},
-                "report": f"Analysis initialization failed: {str(e)}",
-                "saved_image_path": None,
-                "guideline_context": "",
-            }
+            logging.error(f"Pipeline error: {e}", exc_info=True)
+            return f"❌ Error: {e}"