Performance: load a cached Hugging Face embedding model without contacting the Hub

Every chat question and AI suggestion request builds the embedding model, and sentence-transformers
checks the Hub for a newer revision each time. With the model already cached that check cost about
2 seconds per request (0.2 seconds without it), and it fails outright when the Hub is unreachable.

Load from the local cache first and only download when the model is missing. A full rebuild still
checks the Hub, so a new model release is picked up together with re-embedding every document
instead of silently mixing revisions.

docs/administration.md

@@ -565,6 +565,8 @@ Specify `rebuild` to build the index from scratch from all documents in the data
 this the first time you enable the feature, or after changing the embedding backend or
 model.
 
+When using the `huggingface` embedding backend, a model that is already downloaded is loaded from the local cache without contacting the Hugging Face Hub. Only `rebuild` checks the Hub for a newer revision of the model, so a model update is picked up together with re-embedding every document. If the model is not cached yet, it is downloaded on first use.
+
 Specify `update` to incrementally index new and changed documents. This is what the
 scheduled task runs.
 

src/paperless_ai/embedding.py

@@ -1,4 +1,5 @@
 import re
+from contextlib import suppress
 from typing import TYPE_CHECKING
 
 from django.conf import settings
@@ -20,7 +21,17 @@ OCR_LEADER_REGEX = re.compile(r"[._\-\u00b7]{4,}")
 HORIZONTAL_WHITESPACE_REGEX = re.compile(r"[ \t\u00a0]+")
 
 
-def get_embedding_model(config: AIConfig) -> "BaseEmbedding":
+def get_embedding_model(
+    config: AIConfig,
+    *,
+    refresh: bool = False,
+) -> "BaseEmbedding":
+    """
+    Build the embedding model for the configured backend.
+
+    For Hugging Face, a cached model loads offline by default, saving seconds per
+    call. Pass ``refresh=True`` to check the Hub for a newer model revision.
+    """
     match config.llm_embedding_backend:
         case LLMEmbeddingBackend.OPENAI_LIKE:
             from llama_index.embeddings.openai_like import OpenAILikeEmbedding
@@ -52,11 +63,15 @@ def get_embedding_model(config: AIConfig) -> "BaseEmbedding":
         case LLMEmbeddingBackend.HUGGINGFACE:
             from llama_index.embeddings.huggingface import HuggingFaceEmbedding
 
-            return HuggingFaceEmbedding(
-                model_name=config.llm_embedding_model
+            hf_kwargs = {
+                "model_name": config.llm_embedding_model
                 or "sentence-transformers/all-MiniLM-L6-v2",
-                cache_folder=str(settings.DATA_DIR / "hf_cache"),
-            )
+                "cache_folder": str(settings.DATA_DIR / "hf_cache"),
+            }
+            if not refresh:
+                with suppress(OSError):
+                    return HuggingFaceEmbedding(**hf_kwargs, local_files_only=True)
+            return HuggingFaceEmbedding(**hf_kwargs)
         case LLMEmbeddingBackend.OLLAMA:
             from llama_index.embeddings.ollama import OllamaEmbedding
             from ollama import AsyncClient

src/paperless_ai/indexing.py

@@ -470,7 +470,7 @@ def update_llm_index(
         logger.warning("No documents found to index.")
 
     chunk_size = config.llm_embedding_chunk_size
-    embed_model = get_embedding_model(config)
+    embed_model = get_embedding_model(config, refresh=rebuild)
 
     with write_store(embed_model_name=model_name) as store:
         if rebuild or not store.table_exists():

src/paperless_ai/tests/test_ai_indexing.py

@@ -212,6 +212,10 @@ def test_update_llm_index(
 
         ai_config.assert_called_once()
         build_document_node.assert_called_once_with(real_document, chunk_size=512)
+        indexing.get_embedding_model.assert_called_once_with(
+            mock_config,
+            refresh=True,
+        )
 
 
 @pytest.mark.django_db
@@ -336,6 +340,8 @@ def test_update_llm_index_partial_update(
 
         indexing.update_llm_index(rebuild=False)
 
+    assert indexing.get_embedding_model.call_args.kwargs == {"refresh": False}
+
     with indexing.get_vector_store() as store:
         assert store.table_exists(), (
             "Expected the vector store table to exist after incremental update"

src/paperless_ai/tests/test_embedding.py

@@ -1,5 +1,6 @@
 from unittest.mock import ANY
 from unittest.mock import MagicMock
+from unittest.mock import call
 from unittest.mock import patch
 
 import pytest
@@ -167,10 +168,45 @@ def test_get_embedding_model_huggingface(mock_ai_config):
         MockHuggingFaceEmbedding.assert_called_once_with(
             model_name="sentence-transformers/all-MiniLM-L6-v2",
             cache_folder=str(settings.DATA_DIR / "hf_cache"),
+            local_files_only=True,
         )
         assert model == MockHuggingFaceEmbedding.return_value
 
 
+def test_get_embedding_model_huggingface_downloads_when_not_cached(mock_ai_config):
+    mock_ai_config.return_value.llm_embedding_backend = LLMEmbeddingBackend.HUGGINGFACE
+    mock_ai_config.return_value.llm_embedding_model = "some/model"
+    downloaded = MagicMock()
+
+    with patch(
+        "llama_index.embeddings.huggingface.HuggingFaceEmbedding",
+        side_effect=[OSError("not cached"), downloaded],
+    ) as MockHuggingFaceEmbedding:
+        model = get_embedding_model(mock_ai_config.return_value)
+
+    cache_folder = str(settings.DATA_DIR / "hf_cache")
+    assert MockHuggingFaceEmbedding.call_args_list == [
+        call(model_name="some/model", cache_folder=cache_folder, local_files_only=True),
+        call(model_name="some/model", cache_folder=cache_folder),
+    ]
+    assert model is downloaded
+
+
+def test_get_embedding_model_huggingface_refresh_checks_the_hub(mock_ai_config):
+    mock_ai_config.return_value.llm_embedding_backend = LLMEmbeddingBackend.HUGGINGFACE
+    mock_ai_config.return_value.llm_embedding_model = "some/model"
+
+    with patch(
+        "llama_index.embeddings.huggingface.HuggingFaceEmbedding",
+    ) as MockHuggingFaceEmbedding:
+        get_embedding_model(mock_ai_config.return_value, refresh=True)
+
+    MockHuggingFaceEmbedding.assert_called_once_with(
+        model_name="some/model",
+        cache_folder=str(settings.DATA_DIR / "hf_cache"),
+    )
+
+
 def test_get_embedding_model_ollama(mock_ai_config):
     mock_ai_config.return_value.llm_embedding_backend = LLMEmbeddingBackend.OLLAMA
     mock_ai_config.return_value.llm_embedding_model = "embeddinggemma"