🐛 avoid network calls in disconnected environments whn loading fastembed models

Author: m-misiura

nemoguardrails/embeddings/providers/fastembed.py

@@ -14,6 +14,8 @@
 # limitations under the License.
 
 import asyncio
+import os
+from pathlib import Path
 from typing import List
 
 from .base import EmbeddingModel
@@ -49,6 +51,15 @@ def __init__(self, embedding_model: str, **kwargs):
             embedding_model = "sentence-transformers/all-MiniLM-L6-v2"
         self.embedding_model = embedding_model
 
+        # If a pre-downloaded model exists in the cache, pass its path directly
+        # to skip all download attempts. This avoids network calls in disconnected
+        # environments where fastembed would otherwise try HuggingFace Hub and GCS.
+        # See https://github.com/qdrant/fastembed/pull/614
+        if "specific_model_path" not in kwargs:
+            cached_path = self._find_cached_model(embedding_model, Embedding)
+            if cached_path:
+                kwargs["specific_model_path"] = str(cached_path)
+
         try:
             self.model = Embedding(embedding_model, **kwargs)
         except ValueError as ex:
@@ -63,6 +74,34 @@ def __init__(self, embedding_model: str, **kwargs):
         # Get the embedding dimension of the model
         self.embedding_size = len(list(self.model.embed("test"))[0].tolist())
 
+    @staticmethod
+    def _find_cached_model(embedding_model: str, embedding_cls) -> Path | None:
+        """Find a pre-downloaded model in the fastembed cache.
+
+        Looks up the HF source repo for the model in fastembed's registry,
+        then checks if it exists in the HF-convention cache directory
+        (models--{org}--{repo}/snapshots/{hash}/).
+        """
+        cache_dir = Path(os.environ.get("FASTEMBED_CACHE_PATH", str(Path.home() / ".cache" / "fastembed")))
+        if not cache_dir.is_dir():
+            return None
+
+        for model_desc in embedding_cls.list_supported_models():
+            if model_desc["model"] != embedding_model:
+                continue
+            hf_source = (model_desc.get("sources") or {}).get("hf")
+            if not hf_source:
+                break
+            snapshots_dir = cache_dir / f"models--{hf_source.replace('/', '--')}" / "snapshots"
+            if not snapshots_dir.is_dir():
+                break
+            snapshot_dirs = [d for d in snapshots_dir.iterdir() if d.is_dir()]
+            if snapshot_dirs:
+                return snapshot_dirs[0]
+            break
+
+        return None
+
     async def encode_async(self, documents: List[str]) -> List[List[float]]:
         """Encode a list of documents into their corresponding sentence embeddings.