bug fix

Dockerfile

@@ -24,7 +24,19 @@ RUN pip install --no-cache-dir "youtube-transcript-api>=1.2.4"
 
 COPY --chown=appuser:appuser . .
 
-RUN mkdir -p /app/models && python3 -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='ibm-granite/granite-embedding-small-english-r2', local_dir='/app/models/bge-384'); snapshot_download(repo_id='nomic-ai/modernbert-embed-base', local_dir='/app/models/bge-768'); snapshot_download(repo_id='lightonai/modernbert-embed-large', local_dir='/app/models/bge-1024'); snapshot_download(repo_id='nomic-ai/nomic-embed-vision-v1.5', local_dir='/app/models/vision')" && chown -R appuser:appuser /app/models
+RUN mkdir -p /app/models && python3 -c "
+import json, os
+from huggingface_hub import snapshot_download
+snapshot_download(repo_id='ibm-granite/granite-embedding-small-english-r2', local_dir='/app/models/bge-384')
+snapshot_download(repo_id='nomic-ai/modernbert-embed-base', local_dir='/app/models/bge-768')
+snapshot_download(repo_id='lightonai/modernbert-embed-large', local_dir='/app/models/bge-1024')
+snapshot_download(repo_id='nomic-ai/nomic-embed-vision-v1.5', local_dir='/app/models/vision')
+cfg = '/app/models/vision/config.json'
+if os.path.exists(cfg):
+    with open(cfg) as f: d = json.load(f)
+    if isinstance(d.get('n_inner'), float): d['n_inner'] = int(d['n_inner'])
+    with open(cfg, 'w') as f: json.dump(d, f, indent=2)
+" && chown -R appuser:appuser /app/models
 
 RUN mkdir -p /app/logs && \
     chown -R appuser:appuser /app/logs

app/services/embeddings_service.py

@@ -64,6 +64,18 @@ class EmbeddingService:
             return
         local_path = os.path.join(self._models_dir, "vision")
         source = local_path if os.path.isdir(local_path) else _VISION_MODEL_NAME
+
+        cfg_path = os.path.join(local_path, "config.json")
+        if os.path.exists(cfg_path):
+            import json
+            with open(cfg_path) as f:
+                d = json.load(f)
+            if isinstance(d.get("n_inner"), float):
+                d["n_inner"] = int(d["n_inner"])
+                with open(cfg_path, "w") as f:
+                    json.dump(d, f, indent=2)
+                _logger.info("Patched vision model config: n_inner float -> int")
+
         _logger.info("Loading vision embedding model from %s", source)
         self._vision_model = SentenceTransformer(
             source,