Skip to content

Build and run agents you can see, understand and trust.

fix(embedding): reject nonpositive batch sizes (#3451)

PointerLda1fb31a7c63
diff --git a/src/agentscope/embedding/_embedding_base.py b/src/agentscope/embedding/_embedding_base.py
index f5db83843..8f5eb8c03 100644
--- a/src/agentscope/embedding/_embedding_base.py
+++ b/src/agentscope/embedding/_embedding_base.py
@@ -87,7 +87,7 @@ class EmbeddingModelBase(Generic[InputT]):
"""Maximum input length (in tokens) per single input item."""
batch_size: int
- """Maximum number of input items per API call."""
+ """Positive maximum number of input items per API call."""
max_retries: int
"""The maximum number of retries for the underlying API."""
@@ -134,7 +134,7 @@ class EmbeddingModelBase(Generic[InputT]):
context_size (`int`):
Maximum input length (in tokens) per single input item.
batch_size (`int`):
- Maximum number of input items per API call. When
+ Positive maximum number of input items per API call. When
``__call__`` receives more items, it splits them into
batches and calls :meth:`_call_api` concurrently.
max_retries (`int`):
@@ -171,6 +171,10 @@ class EmbeddingModelBase(Generic[InputT]):
raise ValueError(
f"dimensions must be a positive integer, got {dimensions}.",
)
+ if batch_size <= 0:
+ raise ValueError(
+ f"batch_size must be a positive integer, got {batch_size}.",
+ )
self.credential = credential
self.model = model
diff --git a/tests/embedding_base_test.py b/tests/embedding_base_test.py
index 8babba7fd..57716571c 100644
--- a/tests/embedding_base_test.py
+++ b/tests/embedding_base_test.py
@@ -9,7 +9,7 @@ from unittest.mock import AsyncMock, patch
from utils import AnyString
-from agentscope.credential import OpenAICredential
+from agentscope.credential import CredentialBase, OpenAICredential
from agentscope.embedding import (
EmbeddingModelBase,
EmbeddingResponse,
@@ -17,6 +17,34 @@ from agentscope.embedding import (
)
+class _RecordingEmbedding(EmbeddingModelBase[str]):
+ """Offline embedding model that records provider calls."""
+
+ def __init__(self, batch_size: int) -> None:
+ super().__init__(
+ credential=CredentialBase(),
+ model="test",
+ dimensions=1,
+ parameters=None,
+ context_size=10,
+ batch_size=batch_size,
+ max_retries=0,
+ retry_delay=0,
+ )
+ self.calls: list[list[str]] = []
+
+ async def _call_api(
+ self,
+ inputs: list[str],
+ **kwargs: Any,
+ ) -> EmbeddingResponse:
+ self.calls.append(inputs)
+ return EmbeddingResponse(
+ embeddings=[[float(value)] for value in inputs],
+ usage=EmbeddingUsage(tokens=len(inputs), time=0),
+ )
+
+
def _resp(
source: Literal["cache", "api"],
value: float,
@@ -32,7 +60,17 @@ def _resp(
class EmbeddingModelBaseTest(TestCase):
- """Test response merging across cached and API batches."""
+ """Test construction and response merging across batches."""
+
+ def test_batch_size_must_be_positive(self) -> None:
+ """Reject invalid batches before any provider call can be made."""
+ for batch_size in (-1, 0):
+ with self.subTest(batch_size=batch_size):
+ with self.assertRaisesRegex(
+ ValueError,
+ "batch_size must be a positive integer",
+ ):
+ _RecordingEmbedding(batch_size)
def test_merge_single_batch_preserves_cache_source(self) -> None:
"""Preserve the complete response for a single cached batch."""
@@ -124,6 +162,41 @@ class EmbeddingModelBaseTest(TestCase):
)
+class EmbeddingModelBaseBatchingTest(IsolatedAsyncioTestCase):
+ """Exercise valid batch sizes without contacting a provider."""
+
+ async def test_positive_batch_sizes_preserve_inputs_and_embeddings(
+ self,
+ ) -> None:
+ """Small, partial, and oversized batches keep provider input order."""
+ expected_batches = {
+ 1: [["0"], ["1"], ["2"]],
+ 2: [["0", "1"], ["2"]],
+ 5: [["0", "1", "2"]],
+ }
+ for batch_size, batches in expected_batches.items():
+ with self.subTest(batch_size=batch_size):
+ model = _RecordingEmbedding(batch_size)
+
+ response = await model(["0", "1", "2"])
+
+ self.assertEqual(model.calls, batches)
+ self.assertEqual(
+ response.embeddings,
+ [[0.0], [1.0], [2.0]],
+ )
+
+ async def test_empty_inputs_skip_provider_call(self) -> None:
+ """An empty request returns empty vectors and zero usage."""
+ model = _RecordingEmbedding(batch_size=2)
+
+ response = await model([])
+
+ self.assertEqual(model.calls, [])
+ self.assertEqual(response.embeddings, [])
+ self.assertEqual(response.usage, EmbeddingUsage(tokens=0, time=0))
+
+
class _RetryableError(RuntimeError):
"""A provider error eligible for retry."""