Skip to content

Speakr is a personal, self-hosted web application designed for transcribing audio recordings

Python

fix: request float embeddings explicitly and drop the dimensions param on provider rejection instead of failing the pipeline (supersedes #352)

Murtaza Nasirff63ec716d39
diff --git a/src/services/embeddings.py b/src/services/embeddings.py
index 0422aef4..325e14e5 100644
--- a/src/services/embeddings.py
+++ b/src/services/embeddings.py
@@ -137,6 +137,19 @@ def _is_transient_embedding_error(exc):
return any(hint in msg for hint in _TRANSIENT_ERROR_HINTS)
+# Set to True the first time the provider rejects the 'dimensions' parameter
+# (models without Matryoshka support, e.g. BAAI/bge-m3, return a 400). Later
+# calls then omit the parameter instead of failing the whole pipeline. Probing
+# the provider beats a hardcoded list of dimension-capable model names, which
+# would silently disable a setting that newer models do support.
+_dimensions_rejected = False
+
+
+def _is_dimensions_rejection(exc):
+ """True when the error looks like the model rejecting the 'dimensions' kwarg."""
+ return 'dimension' in str(exc).lower()
+
+
def _api_embed(texts, user_id=None):
"""Call the OpenAI-compatible embeddings endpoint and return numpy vectors.
@@ -154,12 +167,17 @@ def _api_embed(texts, user_id=None):
When ``user_id`` is provided and the response includes a ``usage``
block, record the call against the daily token-usage aggregate.
"""
+ global _dimensions_rejected
+
client = get_embedding_api_client()
if client is None or not texts:
return []
- kwargs = {'input': texts, 'model': EMBEDDING_MODEL}
- if EMBEDDING_DIMENSIONS is not None:
+ # encoding_format='float' explicitly: OpenAI SDK v2 defaults to base64,
+ # which some OpenAI-compatible providers cannot produce; float is the
+ # format every compatible provider accepts.
+ kwargs = {'input': texts, 'model': EMBEDDING_MODEL, 'encoding_format': 'float'}
+ if EMBEDDING_DIMENSIONS is not None and not _dimensions_rejected:
kwargs['dimensions'] = EMBEDDING_DIMENSIONS
last_exc = None
@@ -194,6 +212,16 @@ def _api_embed(texts, user_id=None):
except Exception as e:
last_exc = e
+ if 'dimensions' in kwargs and _is_dimensions_rejection(e):
+ _dimensions_rejected = True
+ kwargs.pop('dimensions')
+ current_app.logger.warning(
+ f"Embedding model '{EMBEDDING_MODEL}' rejected the 'dimensions' "
+ f"parameter (EMBEDDING_DIMENSIONS={EMBEDDING_DIMENSIONS}): {e}. "
+ "Retrying without it; the setting will be ignored for the rest "
+ "of this process."
+ )
+ continue
transient = _is_transient_embedding_error(e)
if not transient or attempt == _API_EMBED_MAX_ATTEMPTS:
current_app.logger.error(
diff --git a/tests/test_cov_embeddings.py b/tests/test_cov_embeddings.py
index cf1f46f2..cb1ef439 100644
--- a/tests/test_cov_embeddings.py
+++ b/tests/test_cov_embeddings.py
@@ -377,6 +377,73 @@ def test_api_embed_forwards_dimensions(api_emb):
_restore_env(saved)
+def test_api_embed_sends_float_encoding_format(api_emb):
+ # OpenAI SDK v2 defaults to base64, which some OpenAI-compatible
+ # providers cannot produce; float must be requested explicitly (PR #352).
+ client = _client_returning([[0.1, 0.2]])
+ with app.app_context():
+ with patch.object(api_emb, "get_embedding_api_client", return_value=client):
+ api_emb._api_embed(["x"])
+ assert client.embeddings.create.call_args.kwargs["encoding_format"] == "float"
+
+
+def test_api_embed_drops_dimensions_on_provider_rejection(api_emb):
+ # Models without Matryoshka support (e.g. BAAI/bge-m3) 400 on the
+ # 'dimensions' parameter. The call must retry without it and remember
+ # the rejection for later calls instead of failing the pipeline (PR #352).
+ emb, saved = _reload_embeddings(
+ EMBEDDING_BASE_URL="http://localhost:9999/v1",
+ EMBEDDING_MODEL="bge-m3",
+ EMBEDDING_DIMENSIONS="8",
+ EMBEDDING_API_BACKOFF_SECONDS="0",
+ )
+ try:
+ resp = MagicMock()
+ resp.data = [MagicMock(embedding=[0.0] * 8)]
+ resp.usage = None
+ client = MagicMock()
+ client.embeddings.create.side_effect = [
+ Exception("400: extra_forbidden - dimensions is not supported"),
+ resp,
+ ]
+ with app.app_context():
+ with patch.object(emb, "get_embedding_api_client", return_value=client):
+ out = emb._api_embed(["x"])
+ assert len(out) == 1
+ first, second = client.embeddings.create.call_args_list
+ assert "dimensions" in first.kwargs
+ assert "dimensions" not in second.kwargs
+
+ # A later call must skip dimensions from the start.
+ client2 = _client_returning([[0.0] * 8])
+ with app.app_context():
+ with patch.object(emb, "get_embedding_api_client", return_value=client2):
+ emb._api_embed(["y"])
+ assert "dimensions" not in client2.embeddings.create.call_args.kwargs
+ finally:
+ _restore_env(saved)
+
+
+def test_api_embed_dimensions_kept_when_supported(api_emb):
+ # A provider that accepts dimensions keeps receiving it on every call.
+ emb, saved = _reload_embeddings(
+ EMBEDDING_BASE_URL="http://localhost:9999/v1",
+ EMBEDDING_MODEL="text-embedding-3-large",
+ EMBEDDING_DIMENSIONS="8",
+ EMBEDDING_API_BACKOFF_SECONDS="0",
+ )
+ try:
+ client = _client_returning([[0.0] * 8])
+ with app.app_context():
+ with patch.object(emb, "get_embedding_api_client", return_value=client):
+ emb._api_embed(["x"])
+ emb._api_embed(["y"])
+ for call in client.embeddings.create.call_args_list:
+ assert call.kwargs["dimensions"] == 8
+ finally:
+ _restore_env(saved)
+
+
def test_api_embed_retries_then_succeeds(api_emb):
resp = MagicMock()
resp.data = [MagicMock(embedding=[1.0, 2.0])]