....

app.py

@@ -10,6 +10,7 @@ from typing import Optional, Dict, Any, List, Tuple, Set
 from functools import lru_cache
 import copy
 import re
+from PIL import Image
 from abc import ABC, abstractmethod
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from concurrent.futures import TimeoutError as FuturesTimeoutError
@@ -75,6 +76,18 @@ except ImportError:
     pdfplumber = None
     print("WARNING: pdfplumber library not found. PDF file processing will be unavailable. Install with: pip install pdfplumber")
 
+try:
+    import yt_dlp
+except ImportError:
+    yt_dlp = None
+    print("WARNING: yt-dlp library not found. Video URL processing will be unavailable. Install with: pip install yt-dlp")
+
+try:
+    import cv2
+except ImportError:
+    cv2 = None
+    print("WARNING: opencv-python library not found. Video processing will be unavailable. Install with: pip install opencv-python")
+
 logging.basicConfig(
     level=logging.INFO,
     format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
@@ -96,10 +109,21 @@ MAX_FILE_SIZE = 5 * 1024 * 1024
 CSV_SAMPLE_ROWS = 3
 MAX_FILE_CONTEXT_LENGTH = 10000 
 
+# Global instances for video analysis pipelines
+video_object_detector_pipeline: Optional[Any] = None
+video_species_classifier_pipeline: Optional[Any] = None
+VIDEO_ANALYSIS_DEVICE: int = -1 # -1 for CPU, 0 for CUDA if available and torch is imported
+VIDEO_ANALYSIS_OBJECT_MODEL = "facebook/detr-resnet-50"
+VIDEO_ANALYSIS_SPECIES_MODEL = "openai/clip-vit-base-patch32" # CLIP for zero-shot
+VIDEO_MAX_FRAMES_TO_PROCESS = 120 # Max frames to analyze (e.g., 2 mins at 1fps)
+VIDEO_CONFIDENCE_THRESHOLD_BIRD = 0.6 # Confidence for 'bird' detection
+VIDEO_CONFIDENCE_THRESHOLD_SPECIES = 0.25 # Confidence for species classification via CLIP
+
 asr_pipeline_instance: Optional[Any] = None
 ASR_MODEL_NAME = "openai/whisper-tiny" 
 ASR_PROCESSING_TIMEOUT_SECONDS = 240 
 
+
 DEFAULT_RAG_CONFIG = {
     'search': {
         'tavily_quota': int(os.getenv("TAVILY_QUOTA", "1000")),
@@ -130,6 +154,48 @@ DEFAULT_RAG_CONFIG = {
     'results': {'total_limit': 3, 'enrich_count': 2 }
 }
 
+def _get_video_object_detector():
+    global video_object_detector_pipeline, VIDEO_ANALYSIS_DEVICE
+    if video_object_detector_pipeline is None and hf_transformers_pipeline and torch:
+        try:
+            if VIDEO_ANALYSIS_DEVICE == -1: # Determine device if not already set
+                 if torch.cuda.is_available():
+                    VIDEO_ANALYSIS_DEVICE = 0
+                    gaia_logger.info("CUDA available for video analysis, will use GPU.")
+                 else:
+                    VIDEO_ANALYSIS_DEVICE = -1 # Explicitly CPU
+                    gaia_logger.info("CUDA not available for video analysis, will use CPU.")
+
+            video_object_detector_pipeline = hf_transformers_pipeline(
+                "object-detection",
+                model=VIDEO_ANALYSIS_OBJECT_MODEL,
+                device=VIDEO_ANALYSIS_DEVICE
+            )
+            gaia_logger.info(f"Video Object Detection pipeline ('{VIDEO_ANALYSIS_OBJECT_MODEL}') initialized on {'cuda' if VIDEO_ANALYSIS_DEVICE==0 else 'cpu'}.")
+        except Exception as e:
+            gaia_logger.error(f"Failed to initialize Video Object Detection pipeline: {e}", exc_info=True)
+            return None
+    return video_object_detector_pipeline
+
+def _get_video_species_classifier():
+    global video_species_classifier_pipeline, VIDEO_ANALYSIS_DEVICE
+    if video_species_classifier_pipeline is None and hf_transformers_pipeline and torch:
+        try:
+            if VIDEO_ANALYSIS_DEVICE == -1: # Determine device if not already set (e.g. if detector wasn't called first)
+                 if torch.cuda.is_available(): VIDEO_ANALYSIS_DEVICE = 0
+                 else: VIDEO_ANALYSIS_DEVICE = -1
+
+            video_species_classifier_pipeline = hf_transformers_pipeline(
+                "zero-shot-image-classification",
+                model=VIDEO_ANALYSIS_SPECIES_MODEL,
+                device=VIDEO_ANALYSIS_DEVICE
+            )
+            gaia_logger.info(f"Video Species Classification pipeline ('{VIDEO_ANALYSIS_SPECIES_MODEL}') initialized on {'cuda' if VIDEO_ANALYSIS_DEVICE==0 else 'cpu'}.")
+        except Exception as e:
+            gaia_logger.error(f"Failed to initialize Video Species Classification pipeline: {e}", exc_info=True)
+            return None
+    return video_species_classifier_pipeline
+
 class FileProcessor:
     @staticmethod
     def _get_asr_pipeline():
@@ -540,12 +606,12 @@ class CompositeSearchClient:
         self._def_max_r = self._search_config.get("default_max_results", 3)
     def _init_providers(self, config_dict: Dict) -> List[SearchProvider]:
         providers: List[SearchProvider] = []
-        if TAVILY_API_KEY and TavilyClient:
-            tavily_prov = TavilyProvider(config_dict)
-            if tavily_prov.available(): providers.append(tavily_prov)
         if GOOGLE_CUSTOM_SEARCH_API_KEY and GOOGLE_CUSTOM_SEARCH_CSE_ID:
             google_prov = GoogleProvider(config_dict)
             if google_prov.available(): providers.append(google_prov)
+        if TAVILY_API_KEY and TavilyClient:
+            tavily_prov = TavilyProvider(config_dict)
+            if tavily_prov.available(): providers.append(tavily_prov)
         if DDGS:
             ddgs_prov = DuckDuckGoProvider(config_dict)
             if ddgs_prov.available(): providers.append(ddgs_prov)