Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(ragcon): scope embedding drop_params to official endpoint (#19476)

Sbaaoui Idriss4a1c71baf1cf
diff --git a/rag/llm/cv_model.py b/rag/llm/cv_model.py
index de7659db8e..a29c252209 100644
--- a/rag/llm/cv_model.py
+++ b/rag/llm/cv_model.py
@@ -1355,7 +1355,7 @@ class RAGconCV(GptV4):
RAGcon CV Provider - routes through LiteLLM proxy
Supports vision models through LiteLLM.
- Default Base URL: https://connect.ragcon.ai/v1
+ Default Base URL: https://connect.ragcon.com/v1
"""
_FACTORY_NAME = "RAGcon"
diff --git a/rag/llm/embedding_model.py b/rag/llm/embedding_model.py
index c41cb31d78..cd2a555243 100644
--- a/rag/llm/embedding_model.py
+++ b/rag/llm/embedding_model.py
@@ -267,15 +267,18 @@ class OpenAIEmbed(Base):
self.client = OpenAI(api_key=key, base_url=self.base_url)
self.model_name = model_name
+ def _extra_body(self):
+ return None
+
def _call(self, batch):
- # extra_body is forwarded verbatim to the provider. `drop_params` is
- # an OpenRouter-specific convention; Together AI (and any strict
- # OpenAI-compatible provider) rejects it with HTTP 400
- # "Unrecognized request arguments supplied: drop_params".
# `user` is OpenAI-standard and is only sent when LLM request context
# is active. Local servers that reject unknown fields (LocalAI,
# LM Studio, Xinference) use their own embed classes and omit it.
- res = self.client.embeddings.create(input=batch, model=self.model_name, encoding_format="float", **openai_user_kwargs())
+ kwargs = {"input": batch, "model": self.model_name, "encoding_format": "float", **openai_user_kwargs()}
+ extra_body = self._extra_body()
+ if extra_body is not None:
+ kwargs["extra_body"] = extra_body
+ res = self.client.embeddings.create(**kwargs)
return [d.embedding for d in _sorted_by_index(res.data)], total_token_count_from_response(res)
def encode(self, texts: list):
@@ -1405,7 +1408,7 @@ class RAGconEmbed(OpenAIEmbed):
"""
RAGcon Embedding Provider - routes through LiteLLM proxy
- Default Base URL: https://connect.ragcon.ai/v1
+ Default Base URL: https://connect.ragcon.com/v1
"""
_FACTORY_NAME = "RAGcon"
@@ -1416,6 +1419,12 @@ class RAGconEmbed(OpenAIEmbed):
super().__init__(key, model_name, base_url)
+ def _extra_body(self):
+ host = (urlparse(self.base_url).hostname or "").lower()
+ if host in {"connect.ragcon.com", "connect.ragcon.ai"} or host.endswith((".connect.ragcon.com", ".connect.ragcon.ai")):
+ return {"drop_params": True}
+ return None
+
class PerplexityEmbed(Base):
_FACTORY_NAME = "Perplexity"
diff --git a/rag/llm/tts_model.py b/rag/llm/tts_model.py
index d478205ff0..90fb3e48e2 100644
--- a/rag/llm/tts_model.py
+++ b/rag/llm/tts_model.py
@@ -532,7 +532,7 @@ class RAGconTTS(Base):
RAGcon TTS Provider - routes through LiteLLM proxy
Text-to-speech models routed through LiteLLM.
- Default Base URL: https://connect.ragcon.ai/v1
+ Default Base URL: https://connect.ragcon.com/v1
"""
_FACTORY_NAME = "RAGcon"
diff --git a/test/unit_test/rag/llm/test_embedding_drop_params.py b/test/unit_test/rag/llm/test_embedding_drop_params.py
index 6f9c65bdfe..42af1b4a5c 100644
--- a/test/unit_test/rag/llm/test_embedding_drop_params.py
+++ b/test/unit_test/rag/llm/test_embedding_drop_params.py
@@ -42,7 +42,7 @@ from unittest.mock import MagicMock
import pytest
-from rag.llm.embedding_model import OpenAIEmbed, OpenRouterEmbed
+from rag.llm.embedding_model import OpenAIEmbed, OpenRouterEmbed, RAGconEmbed
# ---------------------------------------------------------------------------
@@ -77,6 +77,13 @@ def _make_openrouter_embed():
return embed
+def _make_ragcon_embed(base_url=None):
+ embed = RAGconEmbed("key", "text-embedding-3-small", base_url=base_url)
+ embed.client = MagicMock()
+ embed.client.embeddings.create = MagicMock(side_effect=lambda input, model, **kwargs: _FakeResp([[float(i)] for i in range(len(input))]))
+ return embed
+
+
# ---------------------------------------------------------------------------
# Tests: OpenAIEmbed (and all its subclasses) must NOT send drop_params
# ---------------------------------------------------------------------------
@@ -162,6 +169,32 @@ def test_openrouter_embed_call_appends_provider_when_provider_order_set():
assert extra_body.get("provider") == {"order": ["Azure", "OpenAI"], "allow_fallbacks": False}
+ "endpoint",
+ [
+ None,
+ "https://connect.ragcon.com/v1",
+ "https://connect.ragcon.ai/v1",
+ "https://api.connect.ragcon.com/v1",
+ "https://api.connect.ragcon.ai/v1",
+ ],
+)
+def test_ragcon_embed_official_endpoint_sends_drop_params(endpoint):
+ embed = _make_ragcon_embed(endpoint)
+ embed._call(["hello"])
+
+ kwargs = embed.client.embeddings.create.call_args.kwargs
+ assert kwargs["extra_body"] == {"drop_params": True}
+
+
+def test_ragcon_embed_custom_endpoint_omits_extra_body():
+ embed = _make_ragcon_embed("https://example.invalid/v1")
+ embed._call(["hello"])
+
+ kwargs = embed.client.embeddings.create.call_args.kwargs
+ assert "extra_body" not in kwargs
+
+
# ---------------------------------------------------------------------------
# Sanity: the response shape is preserved
# ---------------------------------------------------------------------------