fix: make embedding API key optional for local OpenAI-compatible engines (#2261)

Ingesting documents with a local OpenAI-compatible embedding engine (Ollama,
vLLM, ...) failed with `openai.OpenAIError: Missing credentials` when no API
key was configured. The OpenAI-like embedding factory forwarded an empty
api_key straight to the OpenAI client, which rejects empty credentials at
request time.

Chat already tolerates a missing key (completions/openailike.py uses
`... or "default"`) and so does the embedding discovery probe
(`... or "no-key"`); only the embedding factory lacked the fallback, so model
discovery and chat worked while the first real embedding call (ingestion)
crashed.

Mirror the chat behaviour with a placeholder key on the local
OpenAI-compatible path. The strict api.openai.com path is unchanged and still
requires a real key.

Fixes #2260

Co-authored-by: Julian Siegrist <jsiegrist@immomio.de>

private_gpt/components/embedding/factories/openailike.py

@@ -27,7 +27,11 @@ class OpenAILikeEmbeddingFactory(EmbeddingFactory):
         api_base = (
             self.settings.openai.embedding_api_base or self.settings.openai.api_base
         )
-        api_key = self.settings.openai.embedding_api_key or self.settings.openai.api_key
+        api_key = (
+            self.settings.openai.embedding_api_key
+            or self.settings.openai.api_key
+            or "default"
+        )
         model = model_config.name
 
         embedding_model = OpenAILikeEmbedding(

tests/components/embedding/test_openai_embedding_factory.py

@@ -0,0 +1,65 @@
+from unittest.mock import patch
+
+from private_gpt.components.embedding.factories.openai import OpenAIEmbeddingFactory
+from private_gpt.settings.settings import (
+    EmbeddingModelConfig,
+    Settings,
+    unsafe_settings,
+)
+
+
+def _settings(
+    *,
+    api_base: str,
+    api_key: str,
+    embedding_api_base: str | None,
+    embedding_api_key: str | None,
+) -> Settings:
+    settings = Settings(**unsafe_settings)
+    settings.openai.api_base = api_base
+    settings.openai.api_key = api_key
+    settings.openai.embedding_api_base = embedding_api_base
+    settings.openai.embedding_api_key = embedding_api_key
+    return settings
+
+
+def _config() -> EmbeddingModelConfig:
+    return EmbeddingModelConfig(
+        name="mxbai-embed-large", mode="openai", context_window=512
+    )
+
+
+def test_local_openai_compatible_engine_does_not_require_api_key() -> None:
+    # Regression test for #2260: a local engine (Ollama, vLLM, ...) must embed
+    # without an API key instead of failing with "Missing credentials".
+    settings = _settings(
+        api_base="http://localhost:11434/v1",
+        api_key="",
+        embedding_api_base="http://localhost:11434/v1",
+        embedding_api_key=None,
+    )
+
+    with patch(
+        "llama_index.embeddings.openai_like.OpenAILikeEmbedding"
+    ) as mock_embedding:
+        OpenAIEmbeddingFactory(settings)._create_embedding(_config())
+
+    _, kwargs = mock_embedding.call_args
+    assert kwargs["api_key"]
+
+
+def test_real_openai_endpoint_keeps_real_key() -> None:
+    # api.openai.com must keep the real (empty) key and fail loudly instead of
+    # silently receiving a placeholder key.
+    settings = _settings(
+        api_base="https://api.openai.com/v1",
+        api_key="",
+        embedding_api_base=None,
+        embedding_api_key=None,
+    )
+
+    with patch("llama_index.embeddings.openai.OpenAIEmbedding") as mock_embedding:
+        OpenAIEmbeddingFactory(settings)._create_embedding(_config())
+
+    _, kwargs = mock_embedding.call_args
+    assert kwargs["api_key"] == ""