From 159ac03904062953f96cdb9825799a3286f906e4 Mon Sep 17 00:00:00 2001
From: arc53-machine <232052973+arc53-machine@users.noreply.github.com>
Date: Tue, 29 Sep 2026 12:36:57 +0100
Subject: [PATCH] Let FastEmbed download a model when completing its cache
 fails

Completing a partial snapshot is best effort; a network error or rate limit
there is logged and FastEmbed still tries its own sources.
---
 docsgpt/vectorstore/embeddings_local.py    |  8 ++++++--
 tests/vectorstore/test_embeddings_local.py | 11 +++++++++++
 2 files changed, 17 insertions(+), 2 deletions(-)

diff --git a/docsgpt/vectorstore/embeddings_local.py b/docsgpt/vectorstore/embeddings_local.py
index 88573cd5..97cfe939 100644
--- a/docsgpt/vectorstore/embeddings_local.py
+++ b/docsgpt/vectorstore/embeddings_local.py
@@ -263,7 +263,8 @@ def _complete_model_cache(repo: str, cache_dir: Optional[str]) -> None:
     holds. The chunker caches only ``tokenizer.json`` there (see
     ``docsgpt/parser/tokenization.py``), so FastEmbed then finds no graph and
     fails to load. Fetching the files it needs first makes either order work.
-    Offline, nothing is fetched and FastEmbed reports what is missing.
+    Offline, nothing is fetched and FastEmbed reports what is missing; a
+    failed download is logged and left to FastEmbed's own sources too.
 
     Args:
         repo: The model's FastEmbed name.
@@ -293,7 +294,10 @@ def _complete_model_cache(repo: str, cache_dir: Optional[str]) -> None:
     if os.environ.get("HF_HUB_OFFLINE", "").strip().upper() in {"1", "TRUE", "YES", "ON"}:
         return
     logger.warning("The cached %s has no %s; downloading the model files.", source, ", ".join(needed))
-    snapshot_download(repo_id=source, allow_patterns=[*_FASTEMBED_SUPPORT_FILES, *needed], cache_dir=cache)
+    try:
+        snapshot_download(repo_id=source, allow_patterns=[*_FASTEMBED_SUPPORT_FILES, *needed], cache_dir=cache)
+    except Exception as exc:  # noqa: BLE001 -- best effort: FastEmbed still tries its own sources
+        logger.warning("Could not complete the cached %s (%s); leaving the download to FastEmbed.", source, exc)
 
 
 def _pad_to_longest_in_batch(model: Any) -> None:
diff --git a/tests/vectorstore/test_embeddings_local.py b/tests/vectorstore/test_embeddings_local.py
index b3090a1c..f226881b 100644
--- a/tests/vectorstore/test_embeddings_local.py
+++ b/tests/vectorstore/test_embeddings_local.py
@@ -460,6 +460,17 @@ class TestIncompleteModelCache:
         assert "onnx/model.onnx" in downloads[0]["allow_patterns"]
         assert "tokenizer_config.json" in downloads[0]["allow_patterns"]
 
+    def test_a_failed_repair_leaves_loading_to_fastembed(self, monkeypatch, tmp_path):
+        """The repair is best effort: a network error or rate limit here must
+        not stop FastEmbed from trying its own download."""
+        from docsgpt.vectorstore import embeddings_local
+
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising=False)
+        with patch("fastembed.TextEmbedding._list_supported_models", return_value=[self._description()]), \
+                patch("huggingface_hub.hf_hub_download", side_effect=FileNotFoundError("missing")), \
+                patch("huggingface_hub.snapshot_download", side_effect=OSError("rate limited")):
+            embeddings_local._complete_model_cache("sentence-transformers/all-mpnet-base-v2", str(tmp_path))
+
     def test_a_complete_snapshot_downloads_nothing(self, monkeypatch, tmp_path):
         assert self._complete(monkeypatch, tmp_path, cached={"tokenizer.json", "onnx/model.onnx"}) == []