cloud / app.py
dpv007's picture
Update app.py
139d77f verified
"""
Elderly HealthWatch AI Backend (FastAPI) - With GCS Support and Detailed Logging
Enhanced with comprehensive logging at every step
"""
import io
import os
import uuid
import json
import asyncio
import logging
import traceback
import re
from typing import Dict, Any, Optional, Tuple
from datetime import datetime, timedelta
from fastapi import FastAPI, UploadFile, File, BackgroundTasks, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from PIL import Image
import numpy as np
import cv2
# Google Cloud Storage
try:
from google.cloud import storage
from google.oauth2 import service_account
GCS_AVAILABLE = True
except Exception:
GCS_AVAILABLE = False
logging.warning("Google Cloud Storage not available")
try:
from gradio_client import Client, handle_file
GRADIO_AVAILABLE = True
except Exception:
GRADIO_AVAILABLE = False
# ============================================================================
# Configuration
# ============================================================================
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger("elderly_healthwatch")
GRADIO_VLM_SPACE = os.getenv("GRADIO_SPACE", "developer0hye/Qwen3-VL-8B-Instruct")
LLM_GRADIO_SPACE = os.getenv("LLM_GRADIO_SPACE", "Tonic/med-gpt-oss-20b-demo")
HF_TOKEN = os.getenv("HF_TOKEN", None)
# GCS Configuration - Simple!
GCS_BUCKET_NAME = "elderly-healthwatch-images"
GCS_CREDENTIALS_FILE = "gcs-credentials.json"
DEFAULT_VLM_PROMPT = (
"From the provided face/eye images, compute the required screening features "
"(pallor, sclera yellowness, redness, mobility metrics, quality checks) "
"and output a clean JSON feature vector only."
)
LLM_SYSTEM_PROMPT = (
"System: This assistant MUST ONLY OUTPUT a single valid JSON object as its response — "
"no prose, no explanations, no code fences, no annotations."
)
LLM_DEVELOPER_PROMPT = (
"Developer: Output ONLY a single valid JSON object with keys: risk_score, "
"jaundice_probability, anemia_probability, hydration_issue_probability, "
"neurological_issue_probability, summary, recommendation, confidence. "
"Do NOT include any extra fields or natural language outside the JSON object."
)
TMP_DIR = "/tmp/elderly_healthwatch"
os.makedirs(TMP_DIR, exist_ok=True)
# In-memory database
screenings_db: Dict[str, Dict[str, Any]] = {}
# ============================================================================
# Google Cloud Storage Setup
# ============================================================================
def setup_gcs_client():
"""Initialize GCS client from credentials file"""
logger.info("=" * 80)
logger.info("INITIALIZING GOOGLE CLOUD STORAGE")
logger.info("=" * 80)
if not GCS_AVAILABLE:
logger.warning("❌ GCS libraries not installed")
return None, None
try:
logger.info("Looking for credentials file: %s", GCS_CREDENTIALS_FILE)
logger.info("Current working directory: %s", os.getcwd())
logger.info("Credentials file exists: %s", os.path.exists(GCS_CREDENTIALS_FILE))
if os.path.exists(GCS_CREDENTIALS_FILE):
logger.info("✅ Found GCS credentials file")
logger.info("File size: %d bytes", os.path.getsize(GCS_CREDENTIALS_FILE))
credentials = service_account.Credentials.from_service_account_file(GCS_CREDENTIALS_FILE)
logger.info("✅ Credentials loaded successfully")
client = storage.Client(credentials=credentials)
logger.info("✅ Storage client created")
bucket = client.bucket(GCS_BUCKET_NAME)
logger.info("✅ Bucket reference obtained: %s", GCS_BUCKET_NAME)
# Test bucket access
try:
bucket.exists()
logger.info("✅ Bucket access verified")
except Exception as e:
logger.warning("⚠️ Could not verify bucket access: %s", str(e))
logger.info("=" * 80)
logger.info("✅ GCS INITIALIZATION SUCCESSFUL")
logger.info("=" * 80)
return client, bucket
else:
logger.warning("⚠️ GCS credentials file not found at: %s", GCS_CREDENTIALS_FILE)
logger.warning("VLM will use file handles instead of URLs")
return None, None
except Exception as e:
logger.exception("❌ Failed to initialize GCS: %s", str(e))
return None, None
gcs_client, gcs_bucket = setup_gcs_client()
def upload_to_gcs(local_path: str, blob_name: str) -> Optional[str]:
"""Upload file to GCS and return public URL"""
logger.info("-" * 80)
logger.info("UPLOADING TO GCS")
logger.info(" - Local path: %s", local_path)
logger.info(" - Blob name: %s", blob_name)
logger.info(" - File exists: %s", os.path.exists(local_path))
if os.path.exists(local_path):
logger.info(" - File size: %d bytes", os.path.getsize(local_path))
if gcs_bucket is None:
logger.warning(" ❌ GCS bucket not available")
return None
try:
blob = gcs_bucket.blob(blob_name)
logger.info(" - Blob object created")
blob.upload_from_filename(local_path, content_type='image/jpeg')
logger.info(" ✅ File uploaded to GCS")
blob.make_public()
logger.info(" ✅ Blob made public")
public_url = blob.public_url
logger.info(" ✅ Public URL: %s", public_url)
logger.info(" - URL length: %d", len(public_url))
return public_url
except Exception as e:
logger.exception("❌ Failed to upload to GCS: %s", str(e))
return None
# ============================================================================
# Face Detection Setup
# ============================================================================
def setup_face_detector():
"""Initialize face detector (MTCNN or OpenCV fallback)"""
# Try facenet-pytorch MTCNN
try:
from facenet_pytorch import MTCNN
return MTCNN(keep_all=False, device="cpu"), "facenet_pytorch"
except Exception:
pass
# Try classic MTCNN
try:
from mtcnn import MTCNN
return MTCNN(), "mtcnn"
except Exception:
pass
# OpenCV Haar cascade fallback
try:
face_path = os.path.join(cv2.data.haarcascades, "haarcascade_frontalface_default.xml")
eye_path = os.path.join(cv2.data.haarcascades, "haarcascade_eye.xml")
if os.path.exists(face_path) and os.path.exists(eye_path):
return {
"impl": "opencv",
"face_cascade": cv2.CascadeClassifier(face_path),
"eye_cascade": cv2.CascadeClassifier(eye_path)
}, "opencv"
except Exception:
pass
return None, None
face_detector, detector_type = setup_face_detector()
# ============================================================================
# Utility Functions
# ============================================================================
def load_image_from_bytes(bytes_data: bytes) -> Image.Image:
logger.info("Loading image from bytes (size: %d)", len(bytes_data))
img = Image.open(io.BytesIO(bytes_data)).convert("RGB")
logger.info(" - Image loaded: %s, size: %s", img.mode, img.size)
return img
def normalize_probability(val: Optional[float]) -> float:
"""Normalize probability to 0-1 range"""
if val is None:
return 0.0
if val > 1.0 and val <= 100.0:
return max(0.0, min(1.0, val / 100.0))
if val > 100.0:
return 1.0
return max(0.0, min(1.0, val))
def normalize_risk_score(val: Optional[float]) -> float:
"""Normalize risk score to 0-100 range"""
if val is None:
return 0.0
if val <= 1.0:
return round(max(0.0, min(100.0, val * 100.0)), 2)
return round(max(0.0, min(100.0, val)), 2)
# ============================================================================
# Face Detection Functions
# ============================================================================
def detect_face_and_eyes(pil_img: Image.Image) -> Dict[str, Any]:
"""Detect face and eyes, return quality metrics"""
if face_detector is None:
return {
"face_detected": False,
"face_confidence": 0.0,
"eye_openness_score": 0.0,
"left_eye": None,
"right_eye": None
}
img_arr = np.asarray(pil_img)
# Facenet-pytorch MTCNN
if detector_type == "facenet_pytorch":
try:
boxes, probs, landmarks = face_detector.detect(pil_img, landmarks=True)
if boxes is None or len(boxes) == 0:
return {"face_detected": False, "face_confidence": 0.0, "eye_openness_score": 0.0,
"left_eye": None, "right_eye": None}
confidence = float(probs[0]) if probs is not None else 0.0
lm = landmarks[0] if landmarks is not None else None
left_eye = right_eye = None
if lm is not None and len(lm) >= 2:
left_eye = {"x": float(lm[0][0]), "y": float(lm[0][1])}
right_eye = {"x": float(lm[1][0]), "y": float(lm[1][1])}
return {
"face_detected": True,
"face_confidence": confidence,
"eye_openness_score": min(max(confidence * 1.15, 0.0), 1.0),
"left_eye": left_eye,
"right_eye": right_eye
}
except Exception as e:
logger.exception("Facenet MTCNN detection failed")
return {"face_detected": False, "face_confidence": 0.0, "eye_openness_score": 0.0,
"left_eye": None, "right_eye": None}
# Classic MTCNN
elif detector_type == "mtcnn":
try:
detections = face_detector.detect_faces(img_arr)
if not detections:
return {"face_detected": False, "face_confidence": 0.0, "eye_openness_score": 0.0,
"left_eye": None, "right_eye": None}
face = detections[0]
keypoints = face.get("keypoints", {})
confidence = float(face.get("confidence", 0.0))
return {
"face_detected": True,
"face_confidence": confidence,
"eye_openness_score": min(max(confidence * 1.15, 0.0), 1.0),
"left_eye": keypoints.get("left_eye"),
"right_eye": keypoints.get("right_eye")
}
except Exception as e:
logger.exception("Classic MTCNN detection failed")
return {"face_detected": False, "face_confidence": 0.0, "eye_openness_score": 0.0,
"left_eye": None, "right_eye": None}
# OpenCV fallback
elif detector_type == "opencv":
try:
gray = cv2.cvtColor(img_arr, cv2.COLOR_RGB2GRAY)
faces = face_detector["face_cascade"].detectMultiScale(
gray, scaleFactor=1.1, minNeighbors=4, minSize=(60, 60)
)
if len(faces) == 0:
return {"face_detected": False, "face_confidence": 0.0, "eye_openness_score": 0.0,
"left_eye": None, "right_eye": None}
(x, y, w, h) = faces[0]
roi_gray = gray[y:y+h, x:x+w]
eyes = face_detector["eye_cascade"].detectMultiScale(
roi_gray, scaleFactor=1.1, minNeighbors=5, minSize=(20, 10)
)
eye_openness = 1.0 if len(eyes) >= 1 else 0.0
left_eye = None
if len(eyes) >= 1:
ex, ey, ew, eh = eyes[0]
left_eye = {"x": float(x + ex + ew/2), "y": float(y + ey + eh/2)}
confidence = min(1.0, (w*h) / (img_arr.shape[0]*img_arr.shape[1]) * 4.0)
return {
"face_detected": True,
"face_confidence": confidence,
"eye_openness_score": eye_openness,
"left_eye": left_eye,
"right_eye": None
}
except Exception as e:
logger.exception("OpenCV detection failed")
return {"face_detected": False, "face_confidence": 0.0, "eye_openness_score": 0.0,
"left_eye": None, "right_eye": None}
return {"face_detected": False, "face_confidence": 0.0, "eye_openness_score": 0.0,
"left_eye": None, "right_eye": None}
# ============================================================================
# JSON Extraction from LLM Output
# ============================================================================
def extract_json_from_llm_output(raw_text: str) -> Dict[str, Any]:
"""Extract and normalize JSON from LLM output using regex"""
match = re.search(r"\{[\s\S]*\}", raw_text)
if not match:
raise ValueError("No JSON-like block found in LLM output")
block = match.group(0)
def find_number(key: str) -> Optional[float]:
patterns = [
rf'"{key}"\s*:\s*["\']?\s*([-+]?\d+(\.\d+)?)\s*%?\s*["\']?',
rf"'{key}'\s*:\s*['\"]?\s*([-+]?\d+(\.\d+)?)\s*%?\s*['\"]?",
rf'\b{key}\b\s*:\s*["\']?\s*([-+]?\d+(\.\d+)?)\s*%?\s*["\']?',
]
for pat in patterns:
m = re.search(pat, block, flags=re.IGNORECASE)
if m and m.group(1):
try:
return float(m.group(1).replace("%", "").strip())
except Exception:
pass
return None
def find_text(key: str) -> str:
m = re.search(rf'"{key}"\s*:\s*"([^"]*)"', block, flags=re.IGNORECASE)
if m:
return m.group(1).strip()
m = re.search(rf"'{key}'\s*:\s*'([^']*)'", block, flags=re.IGNORECASE)
if m:
return m.group(1).strip()
m = re.search(rf'\b{key}\b\s*:\s*([^\n,}}]+)', block, flags=re.IGNORECASE)
if m:
return m.group(1).strip().strip('",')
return ""
return {
"risk_score": normalize_risk_score(find_number("risk_score")),
"jaundice_probability": round(normalize_probability(find_number("jaundice_probability")), 4),
"anemia_probability": round(normalize_probability(find_number("anemia_probability")), 4),
"hydration_issue_probability": round(normalize_probability(find_number("hydration_issue_probability")), 4),
"neurological_issue_probability": round(normalize_probability(find_number("neurological_issue_probability")), 4),
"confidence": round(normalize_probability(find_number("confidence")), 4),
"summary": find_text("summary"),
"recommendation": find_text("recommendation")
}
# ============================================================================
# VLM Integration - WITH GCS URL SUPPORT AND DETAILED LOGGING
# ============================================================================
def get_gradio_client(space: str) -> Client:
"""Get Gradio client with optional auth"""
if not GRADIO_AVAILABLE:
raise RuntimeError("gradio_client not installed")
return Client(space, hf_token=HF_TOKEN) if HF_TOKEN else Client(space)
def call_vlm_with_urls(face_url: str, eye_url: str, prompt: Optional[str] = None) -> Tuple[Optional[Dict], str]:
"""Call VLM using image URLs instead of file handles"""
logger.info("=" * 80)
logger.info("STARTING VLM CALL WITH URLS")
logger.info("=" * 80)
# Step 1: Validate inputs
logger.info("STEP 1: Validating inputs")
logger.info(" - Face URL provided: %s", bool(face_url))
logger.info(" - Face URL length: %d", len(face_url) if face_url else 0)
logger.info(" - Face URL (full): %s", face_url)
logger.info(" - Eye URL provided: %s", bool(eye_url))
logger.info(" - Eye URL length: %d", len(eye_url) if eye_url else 0)
logger.info(" - Eye URL (full): %s", eye_url)
prompt = prompt or DEFAULT_VLM_PROMPT
logger.info(" - Prompt provided: %s", bool(prompt))
logger.info(" - Prompt length: %d chars", len(prompt))
logger.info(" - Prompt (first 200 chars): %s", prompt[:200])
# Step 2: Initialize Gradio client
logger.info("STEP 2: Initializing Gradio client")
logger.info(" - VLM Space: %s", GRADIO_VLM_SPACE)
logger.info(" - HF Token provided: %s", bool(HF_TOKEN))
try:
client = get_gradio_client(GRADIO_VLM_SPACE)
logger.info(" ✅ Gradio client initialized successfully")
except Exception as e:
logger.error(" ❌ Failed to initialize Gradio client: %s", str(e))
logger.exception("Client initialization error:")
raise
# Step 3: Prepare message formats
logger.info("STEP 3: Preparing message formats")
message_formats = [
# Format 1: URLs in files array
{"text": prompt, "files": [face_url, eye_url]},
# Format 2: Separate image fields
{"prompt": prompt, "image1": face_url, "image2": eye_url},
# Format 3: Single message with URLs
{"message": f"{prompt}\n\nFace image: {face_url}\nEye image: {eye_url}"},
# Format 4: Images array
{"text": prompt, "images": [face_url, eye_url]},
]
logger.info(" - Prepared %d message format variations", len(message_formats))
last_error = None
# Step 4: Try each message format
for idx, message in enumerate(message_formats, 1):
logger.info("-" * 80)
logger.info("STEP 4.%d: Trying message format %d/%d", idx, idx, len(message_formats))
logger.info(" - Format keys: %s", list(message.keys()))
logger.info(" - Format structure (first 300 chars): %s", str(message)[:300])
try:
logger.info(" - Calling VLM API with /chat_fn endpoint...")
result = client.predict(message=message, history=[], api_name="/chat_fn")
logger.info(" ✅ API call succeeded!")
# Step 5: Process result
logger.info("STEP 5: Processing VLM result")
logger.info(" - Result type: %s", type(result))
logger.info(" - Result (first 500 chars): %s", str(result)[:500])
# Parse result structure
if isinstance(result, (list, tuple)):
logger.info(" - Result is list/tuple with %d elements", len(result))
out = result[0] if len(result) > 0 else {}
logger.info(" - Extracted first element, type: %s", type(out))
elif isinstance(result, dict):
logger.info(" - Result is dict with keys: %s", list(result.keys()))
out = result
else:
logger.info(" - Result is %s, converting to dict", type(result))
out = {"text": str(result)}
# Step 6: Extract text from result
logger.info("STEP 6: Extracting text from result")
text_out = None
if isinstance(out, dict):
logger.info(" - Checking dict keys for text content...")
text_out = out.get("text") or out.get("output") or out.get("content") or out.get("response")
logger.info(" - Found text in key: %s",
"text" if out.get("text") else
"output" if out.get("output") else
"content" if out.get("content") else
"response" if out.get("response") else "none")
if not text_out:
logger.info(" - No text in dict, trying alternative methods")
if isinstance(result, str):
text_out = result
logger.info(" - Using result as string directly")
else:
text_out = json.dumps(out)
logger.info(" - Converting to JSON string")
logger.info(" - Extracted text length: %d chars", len(text_out) if text_out else 0)
logger.info(" - Extracted text (first 500 chars): %s", text_out[:500] if text_out else "EMPTY")
logger.info(" - Extracted text (last 200 chars): %s", text_out[-200:] if text_out and len(text_out) > 200 else "")
# Validate extracted text
if not text_out or len(text_out.strip()) == 0:
logger.warning(" ⚠️ VLM returned empty text, trying next format...")
last_error = "Empty response"
continue
# Step 7: Try to parse JSON
logger.info("STEP 7: Attempting to parse JSON from text")
parsed = None
try:
parsed = json.loads(text_out)
if not isinstance(parsed, dict):
logger.warning(" ⚠️ JSON parsed but result is not a dict, type: %s", type(parsed))
parsed = None
else:
logger.info(" ✅ Successfully parsed JSON directly")
logger.info(" - JSON keys: %s", list(parsed.keys()))
logger.info(" - JSON (formatted): %s", json.dumps(parsed, indent=2)[:500])
except Exception as parse_err:
logger.info(" - Direct JSON parsing failed: %s", str(parse_err))
logger.info(" - Attempting to extract JSON from surrounding text...")
# Try to extract JSON from text
try:
first = text_out.find("{")
last = text_out.rfind("}")
logger.info(" - First '{' at position: %d", first)
logger.info(" - Last '}' at position: %d", last)
if first != -1 and last != -1 and last > first:
json_str = text_out[first:last+1]
logger.info(" - Extracted JSON string length: %d", len(json_str))
logger.info(" - Extracted JSON string: %s", json_str[:300])
parsed = json.loads(json_str)
if isinstance(parsed, dict):
logger.info(" ✅ Successfully extracted and parsed JSON from text")
logger.info(" - JSON keys: %s", list(parsed.keys()))
else:
logger.warning(" ⚠️ Extracted JSON is not a dict, type: %s", type(parsed))
parsed = None
else:
logger.warning(" ⚠️ Could not find valid JSON delimiters")
except Exception as extract_err:
logger.warning(" ❌ JSON extraction failed: %s", str(extract_err))
parsed = None
# Step 8: Return successful result
logger.info("=" * 80)
logger.info("✅ VLM CALL COMPLETED SUCCESSFULLY")
logger.info(" - Using message format: %d", idx)
logger.info(" - Parsed JSON: %s", "Yes" if parsed else "No (raw text only)")
logger.info(" - Response length: %d chars", len(text_out))
logger.info("=" * 80)
return parsed, text_out
except Exception as e:
logger.warning(" ❌ VLM format %d failed: %s", idx, str(e))
logger.exception("Detailed error for format %d:", idx)
last_error = str(e)
continue
# All formats failed
logger.error("=" * 80)
logger.error("❌ ALL VLM MESSAGE FORMATS FAILED")
logger.error(" - Tried %d different formats", len(message_formats))
logger.error(" - Last error: %s", last_error)
logger.error("=" * 80)
raise RuntimeError(f"All VLM message formats failed. Last error: {last_error}")
def call_vlm(face_path: str, eye_path: str, prompt: Optional[str] = None) -> Tuple[Optional[Dict], str]:
"""
Call VLM - wrapper that handles both local files and GCS URLs
Strategy: Try GCS first (if available), fallback to file handles
"""
logger.info("=" * 80)
logger.info("VLM CALL ORCHESTRATOR")
logger.info("=" * 80)
logger.info("Input files:")
logger.info(" - Face image: %s", face_path)
logger.info(" - Eye image: %s", eye_path)
logger.info(" - Face exists: %s", os.path.exists(face_path))
logger.info(" - Eye exists: %s", os.path.exists(eye_path))
if os.path.exists(face_path):
logger.info(" - Face size: %d bytes", os.path.getsize(face_path))
if os.path.exists(eye_path):
logger.info(" - Eye size: %d bytes", os.path.getsize(eye_path))
# Strategy 1: Try GCS URLs (if GCS is set up)
if gcs_bucket is not None:
logger.info("🌐 GCS is available, attempting URL-based VLM call")
try:
# Generate unique blob names
unique_id = str(uuid.uuid4())
face_blob_name = f"vlm_temp/{unique_id}_face.jpg"
eye_blob_name = f"vlm_temp/{unique_id}_eye.jpg"
logger.info(" - Generated unique ID: %s", unique_id)
logger.info(" - Face blob name: %s", face_blob_name)
logger.info(" - Eye blob name: %s", eye_blob_name)
# Upload to GCS
logger.info("Uploading images to GCS...")
face_url = upload_to_gcs(face_path, face_blob_name)
eye_url = upload_to_gcs(eye_path, eye_blob_name)
logger.info("Upload results:")
logger.info(" - Face URL: %s", face_url if face_url else "FAILED")
logger.info(" - Eye URL: %s", eye_url if eye_url else "FAILED")
if face_url and eye_url:
logger.info("✅ Successfully uploaded to GCS, calling VLM with URLs")
return call_vlm_with_urls(face_url, eye_url, prompt)
else:
logger.warning("⚠️ GCS upload failed, falling back to file handles")
except Exception as e:
logger.warning("⚠️ GCS error, falling back to file handles: %s", str(e))
logger.exception("GCS error details:")
else:
logger.info("ℹ️ GCS not available, using file handles for VLM")
# Strategy 2: Fallback to file handles (original method)
logger.info("-" * 80)
logger.info("USING FILE HANDLE STRATEGY")
logger.info("-" * 80)
if not os.path.exists(face_path) or not os.path.exists(eye_path):
logger.error("❌ File paths missing!")
logger.error(" - Face exists: %s", os.path.exists(face_path))
logger.error(" - Eye exists: %s", os.path.exists(eye_path))
raise FileNotFoundError("Face or eye image path missing")
prompt = prompt or DEFAULT_VLM_PROMPT
logger.info(" - Using prompt: %s", prompt[:100])
logger.info(" - Creating Gradio client...")
client = get_gradio_client(GRADIO_VLM_SPACE)
logger.info(" ✅ Client created")
logger.info(" - Creating file handles...")
face_handle = handle_file(face_path)
eye_handle = handle_file(eye_path)
logger.info(" ✅ File handles created")
logger.info(" - Face handle type: %s", type(face_handle))
logger.info(" - Eye handle type: %s", type(eye_handle))
message = {"text": prompt, "files": [face_handle, eye_handle]}
logger.info(" - Message structure: %s", list(message.keys()))
try:
logger.info(" - Calling VLM API...")
result = client.predict(message=message, history=[], api_name="/chat_fn")
logger.info(" ✅ VLM API call succeeded")
logger.info(" - Raw result type: %s", type(result))
logger.info(" - Raw result (first 500 chars): %s", str(result)[:500])
except Exception as e:
logger.exception("❌ VLM call with file handles failed")
raise RuntimeError(f"VLM call failed: {e}")
# Process result (same as in call_vlm_with_urls)
logger.info("Processing result...")
if isinstance(result, (list, tuple)):
out = result[0] if len(result) > 0 else {}
logger.info(" - Extracted from list, type: %s", type(out))
elif isinstance(result, dict):
out = result
logger.info(" - Using dict directly")
else:
out = {"text": str(result)}
logger.info(" - Wrapped in dict")
text_out = None
if isinstance(out, dict):
text_out = out.get("text") or out.get("output") or out.get("content")
logger.info(" - Extracted text from dict key")
if not text_out:
if isinstance(result, str):
text_out = result
logger.info(" - Using result string directly")
else:
text_out = json.dumps(out)
logger.info(" - Converted to JSON string")
logger.info(" - Final text length: %d", len(text_out) if text_out else 0)
logger.info(" - Final text (first 500 chars): %s", text_out[:500] if text_out else "EMPTY")
if not text_out or len(text_out.strip()) == 0:
logger.warning(" ⚠️ VLM returned empty text")
text_out = "{}"
parsed = None
try:
parsed = json.loads(text_out)
if not isinstance(parsed, dict):
parsed = None
else:
logger.info(" ✅ Parsed JSON successfully")
except Exception:
logger.info(" - Direct JSON parse failed, trying extraction...")
try:
first = text_out.find("{")
last = text_out.rfind("}")
if first != -1 and last != -1:
parsed = json.loads(text_out[first:last+1])
if not isinstance(parsed, dict):
parsed = None
else:
logger.info(" ✅ Extracted and parsed JSON")
except Exception:
logger.warning(" ⚠️ JSON extraction failed")
pass
logger.info("=" * 80)
logger.info("VLM CALL COMPLETED")
logger.info(" - Parsed JSON: %s", "Yes" if parsed else "No")
logger.info("=" * 80)
return parsed, text_out
# ============================================================================
# LLM Integration
# ============================================================================
def get_fallback_risk_assessment(vlm_output: Any, reason: str = "LLM unavailable") -> Dict[str, Any]:
"""Generate basic risk assessment from VLM output when LLM is unavailable"""
logger.warning("Using fallback risk assessment: %s", reason)
vlm_dict = {}
if isinstance(vlm_output, dict):
vlm_dict = vlm_output
elif isinstance(vlm_output, str):
try:
vlm_dict = json.loads(vlm_output)
except Exception:
pass
has_data = bool(vlm_dict and any(vlm_dict.values()))
if not has_data:
logger.warning("VLM output is empty or invalid, returning minimal assessment")
return {
"risk_score": 0.0,
"jaundice_probability": 0.0,
"anemia_probability": 0.0,
"hydration_issue_probability": 0.0,
"neurological_issue_probability": 0.0,
"confidence": 0.1,
"summary": "Unable to analyze images. Please ensure photos are clear and well-lit.",
"recommendation": "Retake photos with better lighting and clearer view of face and eyes.",
"fallback_mode": True,
"fallback_reason": "no_vlm_data"
}
# Basic heuristic risk scoring
risk_score = 20.0
jaundice_prob = 0.0
anemia_prob = 0.0
hydration_prob = 0.0
neuro_prob = 0.0
sclera_yellow = vlm_dict.get("sclera_yellowness", 0)
pallor = vlm_dict.get("pallor_score", 0)
redness = vlm_dict.get("redness", 0)
if isinstance(sclera_yellow, (int, float)) and sclera_yellow > 0.3:
jaundice_prob = min(0.6, sclera_yellow)
risk_score += 15
if isinstance(pallor, (int, float)) and pallor > 0.4:
anemia_prob = min(0.7, pallor)
risk_score += 20
if isinstance(redness, (int, float)) and redness > 0.5:
hydration_prob = min(0.5, redness * 0.8)
risk_score += 10
return {
"risk_score": round(min(100.0, risk_score), 2),
"jaundice_probability": round(jaundice_prob, 4),
"anemia_probability": round(anemia_prob, 4),
"hydration_issue_probability": round(hydration_prob, 4),
"neurological_issue_probability": round(neuro_prob, 4),
"confidence": 0.4,
"summary": "Basic screening completed. Advanced AI analysis temporarily unavailable.",
"recommendation": "Consider consulting a healthcare professional for a comprehensive assessment.",
"fallback_mode": True,
"fallback_reason": reason
}
def call_llm(vlm_output: Any, use_fallback_on_error: bool = True) -> Dict[str, Any]:
"""Call LLM with VLM output and return structured risk assessment"""
logger.info("=" * 80)
logger.info("STARTING LLM CALL")
logger.info("=" * 80)
if not GRADIO_AVAILABLE:
logger.error("❌ Gradio not available")
if use_fallback_on_error:
return get_fallback_risk_assessment(vlm_output, reason="gradio_not_available")
raise RuntimeError("gradio_client not installed")
vlm_text = vlm_output if isinstance(vlm_output, str) else json.dumps(vlm_output, default=str)
logger.info("VLM input to LLM:")
logger.info(" - Type: %s", type(vlm_output))
logger.info(" - Length: %d chars", len(vlm_text))
logger.info(" - Content (first 500 chars): %s", vlm_text[:500])
if not vlm_text or vlm_text.strip() in ["{}", "[]", ""]:
logger.warning("VLM output is empty, using fallback assessment")
if use_fallback_on_error:
return get_fallback_risk_assessment(vlm_output, reason="empty_vlm_output")
raise RuntimeError("VLM output is empty")
instruction = (
"\n\nSTRICT INSTRUCTIONS:\n"
"1) OUTPUT ONLY a single valid JSON object — no prose, no code fences.\n"
"2) Include keys: risk_score, jaundice_probability, anemia_probability, "
"hydration_issue_probability, neurological_issue_probability, summary, recommendation, confidence.\n"
"3) Use numeric values for probabilities (0-1) and risk_score (0-100).\n"
"4) Use neutral wording in summary/recommendation.\n"
"5) If VLM data is minimal or unclear, set low probabilities and low confidence.\n\n"
"VLM Output:\n" + vlm_text + "\n"
)
logger.info("LLM instruction length: %d chars", len(instruction))
try:
logger.info("Creating LLM client for space: %s", LLM_GRADIO_SPACE)
client = get_gradio_client(LLM_GRADIO_SPACE)
logger.info("✅ LLM client created")
logger.info("Calling LLM with parameters:")
logger.info(" - max_new_tokens: 1024")
logger.info(" - temperature: 0.2")
logger.info(" - top_p: 0.9")
result = client.predict(
input_data=instruction,
max_new_tokens=1024.0,
model_identity=os.getenv("LLM_MODEL_IDENTITY", "GPT-Tonic"),
system_prompt=LLM_SYSTEM_PROMPT,
developer_prompt=LLM_DEVELOPER_PROMPT,
reasoning_effort="medium",
temperature=0.2,
top_p=0.9,
top_k=50,
repetition_penalty=1.0,
api_name="/chat"
)
logger.info("✅ LLM API call succeeded")
logger.info("LLM result type: %s", type(result))
text_out = json.dumps(result) if isinstance(result, (dict, list)) else str(result)
logger.info("LLM raw output length: %d chars", len(text_out))
logger.info("LLM raw output (first 1000 chars):\n%s", text_out[:1000])
logger.info("Attempting to extract JSON from LLM output...")
parsed = extract_json_from_llm_output(text_out)
logger.info("✅ JSON extraction successful")
logger.info("Parsed LLM JSON:\n%s", json.dumps(parsed, indent=2))
# Validation
all_zero = all(
parsed.get(k, 0) == 0
for k in ["jaundice_probability", "anemia_probability",
"hydration_issue_probability", "neurological_issue_probability"]
)
if all_zero and parsed.get("risk_score", 0) == 0:
logger.warning("⚠️ LLM returned all-zero assessment")
parsed["summary"] = "Image analysis incomplete. Please ensure photos are clear and well-lit."
parsed["recommendation"] = "Retake photos with face clearly visible and eyes open."
parsed["confidence"] = 0.1
logger.info("=" * 80)
logger.info("✅ LLM CALL COMPLETED SUCCESSFULLY")
logger.info("=" * 80)
return parsed
except Exception as e:
logger.exception("❌ LLM call failed: %s", str(e))
error_msg = str(e).lower()
if "quota" in error_msg or "gpu" in error_msg:
logger.warning("GPU quota exceeded, using fallback")
if use_fallback_on_error:
return get_fallback_risk_assessment(vlm_output, reason="gpu_quota_exceeded")
if use_fallback_on_error:
return get_fallback_risk_assessment(vlm_output, reason=f"llm_error: {str(e)[:100]}")
raise RuntimeError(f"LLM call failed: {e}")
# ============================================================================
# Background Processing
# ============================================================================
async def process_screening(screening_id: str):
"""Main processing pipeline"""
logger.info("=" * 80)
logger.info("BACKGROUND PROCESSING STARTED FOR: %s", screening_id)
logger.info("=" * 80)
try:
if screening_id not in screenings_db:
logger.error("Screening %s not found in database", screening_id)
return
screenings_db[screening_id]["status"] = "processing"
logger.info("Status updated to: processing")
entry = screenings_db[screening_id]
face_path = entry["face_image_path"]
eye_path = entry["eye_image_path"]
logger.info("Processing images:")
logger.info(" - Face: %s", face_path)
logger.info(" - Eye: %s", eye_path)
# Load images and get quality metrics
logger.info("Loading face image for quality check...")
face_img = Image.open(face_path).convert("RGB")
logger.info(" ✅ Face image loaded: %s", face_img.size)
logger.info("Running face detection...")
detection_result = detect_face_and_eyes(face_img)
logger.info(" - Face detected: %s", detection_result["face_detected"])
logger.info(" - Face confidence: %.3f", detection_result["face_confidence"])
quality_metrics = {
"face_detected": detection_result["face_detected"],
"face_confidence": round(detection_result["face_confidence"], 3),
"face_quality_score": 0.85 if detection_result["face_detected"] else 0.45,
"eye_coords": {
"left_eye": detection_result["left_eye"],
"right_eye": detection_result["right_eye"]
},
"face_brightness": int(np.mean(np.asarray(face_img.convert("L")))),
"face_blur_estimate": int(np.var(np.asarray(face_img.convert("L"))))
}
screenings_db[screening_id]["quality_metrics"] = quality_metrics
logger.info("✅ Quality metrics computed and stored")
# Call VLM
logger.info("Calling VLM...")
vlm_features, vlm_raw = await asyncio.to_thread(call_vlm, face_path, eye_path)
logger.info("✅ VLM call completed")
logger.info(" - Features returned: %s", bool(vlm_features))
logger.info(" - Raw output length: %d", len(vlm_raw) if vlm_raw else 0)
# Call LLM
logger.info("Calling LLM...")
llm_input = vlm_raw if vlm_raw else (vlm_features if vlm_features else "{}")
structured_risk = await asyncio.to_thread(call_llm, llm_input, use_fallback_on_error=True)
logger.info("✅ LLM call completed")
logger.info(" - Risk score: %.2f", structured_risk.get("risk_score", 0))
logger.info(" - Using fallback: %s", structured_risk.get("fallback_mode", False))
# Store results
screenings_db[screening_id]["ai_results"] = {
"vlm_features": vlm_features,
"vlm_raw": vlm_raw,
"structured_risk": structured_risk,
"processing_time_ms": 1200
}
# Build disease predictions
disease_predictions = [
{
"condition": "Anemia-like-signs",
"risk_level": "Medium" if structured_risk["anemia_probability"] > 0.5 else "Low",
"probability": structured_risk["anemia_probability"],
"confidence": structured_risk["confidence"]
},
{
"condition": "Jaundice-like-signs",
"risk_level": "Medium" if structured_risk["jaundice_probability"] > 0.5 else "Low",
"probability": structured_risk["jaundice_probability"],
"confidence": structured_risk["confidence"]
}
]
recommendations = {
"action_needed": "consult" if structured_risk["risk_score"] > 30.0 else "monitor",
"message_english": structured_risk["recommendation"] or "Please follow up with a health professional if concerns persist.",
"message_hindi": ""
}
screenings_db[screening_id].update({
"status": "completed",
"disease_predictions": disease_predictions,
"recommendations": recommendations
})
logger.info("=" * 80)
logger.info("✅ PROCESSING COMPLETED SUCCESSFULLY FOR: %s", screening_id)
logger.info("=" * 80)
except Exception as e:
logger.exception("❌ Processing failed for %s", screening_id)
if screening_id in screenings_db:
screenings_db[screening_id]["status"] = "failed"
screenings_db[screening_id]["error"] = str(e)
# ============================================================================
# FastAPI App & Routes
# ============================================================================
app = FastAPI(title="Elderly HealthWatch AI Backend - Enhanced Logging")
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
@app.get("/")
async def read_root():
return {
"message": "Elderly HealthWatch AI Backend - Enhanced Logging Version",
"gcs_enabled": gcs_bucket is not None,
"version": "1.0.0-gcs-logging"
}
@app.get("/health")
async def health_check():
"""Health check with GCS and LLM status"""
llm_status = "available"
llm_message = None
if GRADIO_AVAILABLE:
try:
client = get_gradio_client(LLM_GRADIO_SPACE)
llm_status = "available"
except Exception as e:
error_msg = str(e).lower()
if "quota" in error_msg or "gpu" in error_msg:
llm_status = "quota_exceeded"
llm_message = "GPU quota exceeded. Using fallback."
else:
llm_status = "error"
llm_message = "LLM temporarily unavailable"
else:
llm_status = "not_installed"
llm_message = "Gradio not available"
return {
"status": "healthy",
"detector": detector_type or "none",
"vlm_available": GRADIO_AVAILABLE,
"vlm_space": GRADIO_VLM_SPACE,
"llm_space": LLM_GRADIO_SPACE,
"llm_status": llm_status,
"llm_message": llm_message,
"gcs_available": gcs_bucket is not None,
"gcs_bucket": GCS_BUCKET_NAME if gcs_bucket else None,
"fallback_enabled": True,
"enhanced_logging": True
}
@app.post("/api/v1/validate-eye-photo")
async def validate_eye_photo(image: UploadFile = File(...)):
"""Validate eye photo quality"""
if face_detector is None:
raise HTTPException(status_code=500, detail="No face detector available")
try:
content = await image.read()
if not content:
raise HTTPException(status_code=400, detail="Empty file")
pil_img = load_image_from_bytes(content)
result = detect_face_and_eyes(pil_img)
if not result["face_detected"]:
return {
"valid": False,
"face_detected": False,
"eye_openness_score": 0.0,
"message_english": "No face detected. Please ensure your face is clearly visible.",
"message_hindi": "कोई चेहरा नहीं मिला। कृपया सुनिश्चित करें कि आपका चेहरा स्पष्ट रूप से दिखाई दे।"
}
is_valid = result["eye_openness_score"] >= 0.3
return {
"valid": is_valid,
"face_detected": True,
"eye_openness_score": round(result["eye_openness_score"], 2),
"message_english": "Photo looks good! Eyes are properly open." if is_valid
else "Eyes appear closed. Please open your eyes wide and try again.",
"message_hindi": "फोटो अच्छी है! आंखें ठीक से खुली हैं।" if is_valid
else "आंखें बंद दिखाई दे रही हैं। कृपया अपनी आंखें चौड़ी खोलें।",
"eye_landmarks": {
"left_eye": result["left_eye"],
"right_eye": result["right_eye"]
}
}
except Exception as e:
logger.exception("Validation failed")
raise HTTPException(status_code=500, detail=str(e))
@app.post("/api/v1/upload")
async def upload_images(
background_tasks: BackgroundTasks,
face_image: UploadFile = File(...),
eye_image: UploadFile = File(...)
):
"""Upload images and start background processing"""
logger.info("=" * 80)
logger.info("NEW UPLOAD REQUEST")
logger.info("=" * 80)
try:
screening_id = str(uuid.uuid4())
logger.info("Generated screening ID: %s", screening_id)
face_path = os.path.join(TMP_DIR, f"{screening_id}_face.jpg")
eye_path = os.path.join(TMP_DIR, f"{screening_id}_eye.jpg")
logger.info("Reading uploaded files...")
face_bytes = await face_image.read()
eye_bytes = await eye_image.read()
logger.info(" - Face image: %d bytes", len(face_bytes))
logger.info(" - Eye image: %d bytes", len(eye_bytes))
logger.info("Saving images to disk...")
with open(face_path, "wb") as f:
f.write(face_bytes)
with open(eye_path, "wb") as f:
f.write(eye_bytes)
logger.info(" ✅ Images saved")
screenings_db[screening_id] = {
"id": screening_id,
"timestamp": datetime.utcnow().isoformat() + "Z",
"face_image_path": face_path,
"eye_image_path": eye_path,
"status": "queued",
"quality_metrics": {},
"ai_results": {},
"disease_predictions": [],
"recommendations": {}
}
logger.info("Adding background task...")
background_tasks.add_task(process_screening, screening_id)
logger.info("✅ Upload successful, processing queued")
return {"screening_id": screening_id}
except Exception as e:
logger.exception("❌ Upload failed")
raise HTTPException(status_code=500, detail=str(e))
@app.post("/api/v1/analyze/{screening_id}")
async def analyze_screening(screening_id: str, background_tasks: BackgroundTasks):
"""Re-analyze existing screening"""
if screening_id not in screenings_db:
raise HTTPException(status_code=404, detail="Screening not found")
if screenings_db[screening_id]["status"] == "processing":
return {"message": "Already processing"}
screenings_db[screening_id]["status"] = "queued"
background_tasks.add_task(process_screening, screening_id)
return {"message": "Analysis enqueued"}
@app.get("/api/v1/status/{screening_id}")
async def get_status(screening_id: str):
"""Get processing status"""
if screening_id not in screenings_db:
raise HTTPException(status_code=404, detail="Screening not found")
status = screenings_db[screening_id]["status"]
progress = 50 if status == "processing" else (100 if status == "completed" else 0)
return {"screening_id": screening_id, "status": status, "progress": progress}
@app.get("/api/v1/results/{screening_id}")
async def get_results(screening_id: str):
"""Get screening results"""
if screening_id not in screenings_db:
raise HTTPException(status_code=404, detail="Screening not found")
return screenings_db[screening_id]
@app.get("/api/v1/history/{user_id}")
async def get_history(user_id: str):
"""Get user screening history"""
history = [s for s in screenings_db.values() if s.get("user_id") == user_id]
return {"screenings": history}
@app.post("/api/v1/get-vitals")
async def get_vitals_from_upload(
face_image: UploadFile = File(...),
eye_image: UploadFile = File(...)
):
"""Synchronous VLM + LLM pipeline with GCS support"""
logger.info("=" * 80)
logger.info("GET VITALS REQUEST (SYNCHRONOUS)")
logger.info("=" * 80)
if not GRADIO_AVAILABLE:
raise HTTPException(
status_code=503,
detail="AI services temporarily unavailable."
)
try:
uid = str(uuid.uuid4())
face_path = os.path.join(TMP_DIR, f"{uid}_face.jpg")
eye_path = os.path.join(TMP_DIR, f"{uid}_eye.jpg")
logger.info("Reading and saving images...")
face_bytes = await face_image.read()
eye_bytes = await eye_image.read()
with open(face_path, "wb") as f:
f.write(face_bytes)
with open(eye_path, "wb") as f:
f.write(eye_bytes)
logger.info("✅ Images saved: %d + %d bytes", len(face_bytes), len(eye_bytes))
# Call VLM
vlm_features, vlm_raw = await asyncio.to_thread(call_vlm, face_path, eye_path)
# Call LLM
llm_input = vlm_raw if vlm_raw else (vlm_features if vlm_features else "{}")
structured_risk = await asyncio.to_thread(call_llm, llm_input, use_fallback_on_error=True)
logger.info("=" * 80)
logger.info("✅ GET VITALS COMPLETED")
logger.info("=" * 80)
return {
"vlm_features": vlm_features,
"vlm_raw": vlm_raw,
"structured_risk": structured_risk,
"using_fallback": structured_risk.get("fallback_mode", False),
"using_gcs": gcs_bucket is not None
}
except Exception as e:
logger.exception("❌ Get vitals failed")
error_msg = str(e).lower()
if "quota" in error_msg or "gpu" in error_msg:
raise HTTPException(
status_code=503,
detail="AI service at capacity. Please try again in a few minutes."
)
raise HTTPException(
status_code=500,
detail="Unable to process images. Please try again."
)
@app.post("/api/v1/get-vitals/{screening_id}")
async def get_vitals_for_screening(screening_id: str):
"""Re-run VLM + LLM on existing screening"""
logger.info("GET VITALS FOR EXISTING SCREENING: %s", screening_id)
if screening_id not in screenings_db:
raise HTTPException(status_code=404, detail="Screening not found")
entry = screenings_db[screening_id]
face_path = entry.get("face_image_path")
eye_path = entry.get("eye_image_path")
if not (face_path and os.path.exists(face_path) and eye_path and os.path.exists(eye_path)):
raise HTTPException(status_code=400, detail="Images missing")
try:
vlm_features, vlm_raw = await asyncio.to_thread(call_vlm, face_path, eye_path)
llm_input = vlm_raw if vlm_raw else (vlm_features if vlm_features else "{}")
structured_risk = await asyncio.to_thread(call_llm, llm_input, use_fallback_on_error=True)
entry.setdefault("ai_results", {}).update({
"vlm_features": vlm_features,
"vlm_raw": vlm_raw,
"structured_risk": structured_risk,
"last_vitals_run": datetime.utcnow().isoformat() + "Z",
"using_fallback": structured_risk.get("fallback_mode", False)
})
logger.info("✅ Get vitals completed for screening: %s", screening_id)
return {
"screening_id": screening_id,
"vlm_features": vlm_features,
"vlm_raw": vlm_raw,
"structured_risk": structured_risk,
"using_fallback": structured_risk.get("fallback_mode", False)
}
except Exception as e:
logger.exception("❌ Get vitals for screening failed")
raise HTTPException(status_code=500, detail=str(e))
if __name__ == "__main__":
import uvicorn
logger.info("=" * 80)
logger.info("STARTING FASTAPI SERVER")
logger.info("=" * 80)
uvicorn.run("app_with_detailed_logs:app", host="0.0.0.0", port=7860, reload=False)