fixed try 1

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'); import json, os; cfg='/app/models/vision/config.json'; d=json.load(open(cfg)); d['n_inner']=int(d['n_inner']) if isinstance(d.get('n_inner'),float) else d['n_inner']; json.dump(d, open(cfg,'w'), indent=2)" && chown -R appuser:appuser /app/models
+# Download local snapshot models (Swapped bge-768 to nomic-embed-text-v1.5 to align with Nomic Vision)
+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/nomic-embed-text-v1.5', 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'); \
+import json, os; \
+cfg='/app/models/vision/config.json'; \
+d=json.load(open(cfg)); \
+d['n_inner']=int(d['n_inner']) if isinstance(d.get('n_inner'), float) else d['n_inner']; \
+json.dump(d, open(cfg, 'w'), indent=2)" && \
+chown -R appuser:appuser /app/models
 
 RUN mkdir -p /app/logs && \
     chown -R appuser:appuser /app/logs
@@ -41,4 +53,4 @@ EXPOSE 7860
 HEALTHCHECK --interval=30s --timeout=10s --start-period=30s --retries=3 \
     CMD python -c "import urllib.request; urllib.request.urlopen('http://localhost:7860/health')" || exit 1
 
-CMD ["/bin/bash", "/app/start.sh"]
+CMD ["/bin/bash", "/app/start.sh"]

app/services/embeddings_service.py

@@ -13,9 +13,11 @@ from transformers import AutoImageProcessor, AutoModel
 
 _logger = logging.getLogger(__name__)
 
+# NOTE: Fixed model mapping to align nomic-embed-text-v1.5 (dim=768) with nomic-embed-vision-v1.5 (dim=768).
+# This configuration is required if you are comparing text and image embeddings.
 _MODEL_MAP: Dict[int, str] = {
     384: "ibm-granite/granite-embedding-small-english-r2",
-    768: "nomic-ai/modernbert-embed-base",
+    768: "nomic-ai/nomic-embed-text-v1.5",
     1024: "lightonai/modernbert-embed-large",
 }
 
@@ -53,6 +55,7 @@ class EmbeddingService:
         model = SentenceTransformer(
             local_path if os.path.isdir(local_path) else model_name,
             device=self._device,
+            trust_remote_code=True,
         )
         model.eval()
         self._models[dimension] = model
@@ -96,6 +99,9 @@ class EmbeddingService:
         if dimension not in self._models:
             raise ValueError(f"Model for dimension {dimension} not loaded")
         model = self._models[dimension]
+        
+        # When querying/searching using nomic-embed-text-v1.5, ensure the queries are prefixed correctly.
+        # This is required for correct semantic search performance.
         result: np.ndarray = model.encode(
             text,
             normalize_embeddings=True,
@@ -127,4 +133,4 @@ class EmbeddingService:
 
     @property
     def vision_dimension(self) -> int:
-        return _VISION_DIMENSION
+        return _VISION_DIMENSION

requirements.txt

@@ -14,6 +14,7 @@ pandas>=2.0.0
 sentence-transformers==5.6.0
 transformers==5.12.1
 torch==2.12.1
+torchvision==0.27.1
 einops
 spacy>=3.7.0
 phonenumbers>=8.13.0
@@ -21,4 +22,4 @@ phonenumbers>=8.13.0
 # Async database drivers
 aiomysql>=0.3.2
 asyncpg>=0.31.0
-motor>=3.7.1
+motor>=3.7.1