Skip to content

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

fix(llm): honor the batch size BaiduYiyanEmbed already declares (#19690)

Lukas Buck9f31a2459ea2
diff --git a/rag/llm/embedding_model.py b/rag/llm/embedding_model.py
index f338ac32e4..4efc832164 100644
--- a/rag/llm/embedding_model.py
+++ b/rag/llm/embedding_model.py
@@ -1187,27 +1187,20 @@ class BaiduYiyanEmbed(Base):
self.client = qianfan.Embedding(access_token=key_obj)
self.model_name = model_name
+ def _call(self, batch):
+ res = self.client.do(model=self.model_name, texts=batch).body
+ return [r["embedding"] for r in res["data"]], total_token_count_from_response(res)
+
def encode(self, texts: list, batch_size=16):
- try:
- res = self.client.do(model=self.model_name, texts=texts).body
- return (
- np.array([r["embedding"] for r in res["data"]]),
- total_token_count_from_response(res),
- )
- except Exception as _e:
- logger.exception("BaiduYiyanEmbed: embedding request failed")
- raise EmbeddingError(f"Embedding request failed for BaiduYiyanEmbed. Error: {_e}") from _e
+ # `batch_size` has been part of this signature since the class was added but the
+ # request went out whole, so a document with more chunks than the provider accepts
+ # per call failed as a whole. Drive the shared template instead, as every other
+ # OpenAI-style provider here does.
+ return self._batched_encode(texts, self._call, batch_size=batch_size)
def encode_queries(self, text):
- try:
- res = self.client.do(model=self.model_name, texts=[text]).body
- return (
- np.array(res["data"][0]["embedding"]),
- total_token_count_from_response(res),
- )
- except Exception as _e:
- logger.exception("BaiduYiyanEmbed: query embedding request failed")
- raise EmbeddingError(f"Embedding request failed for BaiduYiyanEmbed. Error: {_e}") from _e
+ vectors, token_count = self._batched_encode([text], self._call, batch_size=1)
+ return vectors[0], token_count
class VoyageEmbed(Base):
diff --git a/test/unit_test/rag/llm/test_embedding_model.py b/test/unit_test/rag/llm/test_embedding_model.py
index 4b32ee8583..087efa5c72 100644
--- a/test/unit_test/rag/llm/test_embedding_model.py
+++ b/test/unit_test/rag/llm/test_embedding_model.py
@@ -309,6 +309,49 @@ class TestBatching:
vectors, _ = embed.encode(["a", "b", "c"])
np.testing.assert_array_equal(vectors[:, 0], np.array([0.0, 1.0, 2.0]))
+ def test_baidu_yiyan_issues_ceil_n_over_batch_calls(self):
+ """`batch_size` was in the signature and unused: the whole document went in one call."""
+ embed = BaiduYiyanEmbed.__new__(BaiduYiyanEmbed)
+ embed.model_name = "bge-large-zh"
+ embed.client = MagicMock()
+ embed.client.do = MagicMock(side_effect=lambda model, texts: SimpleNamespace(body={"data": [{"embedding": [float(len(t))]} for t in texts], "usage": {"total_tokens": len(texts)}}))
+
+ texts = [f"t{i}" for i in range(40)] # batch_size 16 -> ceil(40/16) == 3
+ vectors, token_count = embed.encode(texts)
+
+ assert embed.client.do.call_count == 3
+ assert [len(call.kwargs["texts"]) for call in embed.client.do.call_args_list] == [16, 16, 8]
+ assert vectors.shape == (40, 1)
+ assert token_count == 40
+
+ def test_baidu_yiyan_preserves_order_across_batches(self):
+ embed = BaiduYiyanEmbed.__new__(BaiduYiyanEmbed)
+ embed.model_name = "bge-large-zh"
+ embed.client = MagicMock()
+ embed.client.do = MagicMock(side_effect=lambda model, texts: SimpleNamespace(body={"data": [{"embedding": [float(len(t))]} for t in texts], "usage": {"total_tokens": 1}}))
+
+ texts = ["a" * (i + 1) for i in range(20)] # spans two batches
+ vectors, _ = embed.encode(texts)
+
+ np.testing.assert_array_equal(vectors[:, 0], np.array([float(len(t)) for t in texts]))
+
+ def test_baidu_yiyan_surfaces_a_failed_batch(self):
+ embed = BaiduYiyanEmbed.__new__(BaiduYiyanEmbed)
+ embed.model_name = "bge-large-zh"
+ embed.client = MagicMock()
+ calls = {"n": 0}
+
+ def _do(model, texts):
+ calls["n"] += 1
+ if calls["n"] == 2:
+ raise RuntimeError("batch too large")
+ return SimpleNamespace(body={"data": [{"embedding": [0.0]} for _ in texts], "usage": {"total_tokens": 1}})
+
+ embed.client.do = MagicMock(side_effect=_do)
+
+ with pytest.raises(EmbeddingError, match="Embedding request failed for BaiduYiyanEmbed"):
+ embed.encode([f"t{i}" for i in range(40)])
+
def test_ollama_issues_ceil_n_over_batch_calls(self):
embed = OllamaEmbed("x", "nomic-embed-text", base_url="http://localhost:11434")
embed.client = MagicMock()