Move model download to runtime with auth

Dockerfile

@@ -20,7 +20,6 @@ RUN pip install --no-cache-dir -r requirements.txt
 # This ensures it gets the exact file without guessing the version
 RUN pip install https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.7.1/en_core_web_sm-3.7.1.tar.gz
 
-RUN python -c "from fast_gliner import FastGLiNER; FastGLiNER.from_pretrained('Xenova/gliner-small-v2.1')"
 
 COPY . .
 

app/services/ai_extractor.py

@@ -1,41 +1,52 @@
-# 1. Change Import
+import os
 from fast_gliner import FastGLiNER
 
 class UniversalExtractor:
     def __init__(self):
         print("Loading Fast GLiNER model (ONNX)...")
-        # 2. Use the ONNX version of the same model
-        # "Xenova/gliner-small-v2.1" is the optimized version of "urchade/gliner_small-v2.1"
-        self.model = FastGLiNER.from_pretrained("Xenova/gliner-small-v2.1")
+        
+        # 1. Get the token securely from Hugging Face Secrets
+        hf_token = os.getenv("HF_TOKEN")
+        
+        # 2. Pass the token explicitly. 
+        # If token is None, it tries anonymous (which might fail in Spaces).
+        # If token is valid, it authenticates and succeeds.
+        self.model = FastGLiNER.from_pretrained(
+            "Xenova/gliner-small-v2.1", 
+            token=hf_token
+        )
         
         self.labels = [
-            "Skill", "Software",]
+            "Skill", "Software",
+        ]
 
     def extract_universal(self, text: str):
-        # 3. Use batch processing for extra speed
-        # We split text into chunks and process them in parallel
+        # Limit text length to prevent memory spikes
         chunks = [text[i:i+3000] for i in range(0, len(text), 3000)]
         
         extracted_data = {
             "skills": set(),
             "tools": set(),
+          
         }
         
-        # FastGLiNER's batch_predict is highly optimized
-        # We pass all chunks at once
         if chunks:
-            batch_results = self.model.batch_predict_entities(chunks, self.labels, threshold=0.35)
-            
-            for entities in batch_results:
-                for entity in entities:
-                    text_val = entity["text"].strip()
-                    label = entity["label"]
-                    
-                    if label == "Skill":
-                        extracted_data["skills"].add(text_val)
-                    elif label == "Software":
-                        extracted_data["tools"].add(text_val)
-                    
+            # Batch predict
+            try:
+                batch_results = self.model.batch_predict_entities(chunks, self.labels, threshold=0.35)
+                
+                for entities in batch_results:
+                    for entity in entities:
+                        text_val = entity["text"].strip()
+                        label = entity["label"]
+                        
+                        if label == "Skill":
+                            extracted_data["skills"].add(text_val)
+                        elif label == "Software":
+                            extracted_data["tools"].add(text_val)
+                        
+            except Exception as e:
+                print(f"GLiNER Prediction Error: {e}")
 
         return {k: sorted(list(v)) for k, v in extracted_data.items()}