Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(embedding): support stable Gemini Embedding 2 (#6383)

z1gonb5a389245661
diff --git a/docs/en/guides/01-configuration.md b/docs/en/guides/01-configuration.md
index 2c7319612..eceac1ec2 100644
--- a/docs/en/guides/01-configuration.md
+++ b/docs/en/guides/01-configuration.md
@@ -573,7 +573,7 @@ Version 1.39.0 is the minimum SDK version with the client context managers and c
"dense": {
"provider": "gemini",
"api_key": "your-google-api-key",
- "model": "gemini-embedding-2-preview",
+ "model": "gemini-embedding-2",
"dimension": 3072
}
}
@@ -581,12 +581,18 @@ Version 1.39.0 is the minimum SDK version with the client context managers and c
```
Available Gemini embedding models:
-- `gemini-embedding-2-preview`: 8192 token input limit, 1–3072 output dimension (MRL)
+- `gemini-embedding-2`: 8192 token input limit, 1–3072 output dimension (MRL)
- `gemini-embedding-001`: 2048 token input limit, 1–3072 output dimension (MRL)
- `text-embedding-004`: 2048 token input limit, 768 output dimension (fixed)
Recommended dimensions: `768`, `1536`, or `3072` (default: `3072`).
+Gemini Embedding 2 represents task types as documented text instructions because
+the stable model does not accept the `task_type` request field. OpenViking applies
+these instructions when `query_param` or `document_param` is configured. If you
+change the model for an existing index, rebuild its vectors unless the provider
+explicitly guarantees that the two embedding spaces are compatible.
+
Get your API key at https://aistudio.google.com/apikey
**DashScope (Alibaba Tongyi) provider:**
@@ -648,7 +654,7 @@ Get your API key at https://dashscope.console.aliyun.com/api-key
"dense": {
"provider": "gemini",
"api_key": "your-google-api-key",
- "model": "gemini-embedding-2-preview",
+ "model": "gemini-embedding-2",
"dimension": 3072,
"query_param": "RETRIEVAL_QUERY",
"document_param": "RETRIEVAL_DOCUMENT"
diff --git a/docs/zh/guides/01-configuration.md b/docs/zh/guides/01-configuration.md
index 5822ac572..4aac9c79c 100644
--- a/docs/zh/guides/01-configuration.md
+++ b/docs/zh/guides/01-configuration.md
@@ -570,7 +570,7 @@ SDK 1.39.0 起提供所需的客户端上下文管理器和关闭方法,用于
"dense": {
"provider": "gemini",
"api_key": "your-google-api-key",
- "model": "gemini-embedding-2-preview",
+ "model": "gemini-embedding-2",
"dimension": 3072
}
}
@@ -578,12 +578,16 @@ SDK 1.39.0 起提供所需的客户端上下文管理器和关闭方法,用于
```
可用 Gemini 嵌入模型:
-- `gemini-embedding-2-preview`: 8192 token 输入限制, 1–3072 输出维度 (MRL)
+- `gemini-embedding-2`: 8192 token 输入限制, 1–3072 输出维度 (MRL)
- `gemini-embedding-001`: 2048 token 输入限制, 1–3072 输出维度 (MRL)
- `text-embedding-004`: 2048 token 输入限制, 768 输出维度(固定)
推荐维度: `768`、`1536` 或 `3072`(默认: `3072`)。
+稳定版 Gemini Embedding 2 不接受 `task_type` 请求字段,而是使用文档规定的文本任务指令。
+配置 `query_param` 或 `document_param` 后,OpenViking 会自动添加相应指令。如果修改已有索引的
+embedding 模型,除非 provider 明确保证两个向量空间兼容,否则应重新生成该索引的向量。
+
获取 API Key: https://aistudio.google.com/apikey
**DashScope(阿里通义)provider 配置示例:**
@@ -645,7 +649,7 @@ SDK 1.39.0 起提供所需的客户端上下文管理器和关闭方法,用于
"dense": {
"provider": "gemini",
"api_key": "your-google-api-key",
- "model": "gemini-embedding-2-preview",
+ "model": "gemini-embedding-2",
"dimension": 3072,
"query_param": "RETRIEVAL_QUERY",
"document_param": "RETRIEVAL_DOCUMENT"
diff --git a/openviking/models/embedder/gemini_embedders.py b/openviking/models/embedder/gemini_embedders.py
index 1963edb58..3c44afeb1 100644
--- a/openviking/models/embedder/gemini_embedders.py
+++ b/openviking/models/embedder/gemini_embedders.py
@@ -26,33 +26,35 @@ from openviking_cli.utils import get_logger
logger = get_logger(__name__)
# Keep for backward-compat with existing unit tests that import it
-_GEMINI_INPUT_TOKEN_LIMIT = 8192 # gemini-embedding-2-preview hard limit
+_GEMINI_INPUT_TOKEN_LIMIT = 8192 # Gemini Embedding 2 hard limit
# Per-model token limits (Google API hard limits, from official docs)
_MODEL_TOKEN_LIMITS: Dict[str, int] = {
+ "gemini-embedding-2": 8192,
"gemini-embedding-2-preview": 8192,
"gemini-embedding-001": 2048,
}
_DEFAULT_TOKEN_LIMIT = 2048 # conservative fallback for unknown future models
-_VALID_TASK_TYPES: frozenset = frozenset(
- {
- "RETRIEVAL_QUERY",
- "RETRIEVAL_DOCUMENT",
- "SEMANTIC_SIMILARITY",
- "CLASSIFICATION",
- "CLUSTERING",
- "QUESTION_ANSWERING",
- "FACT_VERIFICATION",
- "CODE_RETRIEVAL_QUERY",
- }
-)
+# Stable Embedding 2 uses text instructions instead of task_type/title fields.
+# https://ai.google.dev/gemini-api/docs/embeddings#task-types
+_TASK_INSTRUCTIONS: Dict[str, str] = {
+ "RETRIEVAL_QUERY": "task: search result | query: {text}",
+ "RETRIEVAL_DOCUMENT": "title: {title} | text: {text}",
+ "SEMANTIC_SIMILARITY": "task: sentence similarity | query: {text}",
+ "CLASSIFICATION": "task: classification | query: {text}",
+ "CLUSTERING": "task: clustering | query: {text}",
+ "QUESTION_ANSWERING": "task: question answering | query: {text}",
+ "FACT_VERIFICATION": "task: fact checking | query: {text}",
+ "CODE_RETRIEVAL_QUERY": "task: code retrieval | query: {text}",
+}
+_VALID_TASK_TYPES: frozenset = frozenset(_TASK_INSTRUCTIONS)
_ERROR_HINTS: Dict[int, str] = {
400: "Invalid request — check model name and task_type value.",
401: "Invalid API key. Verify your GOOGLE_API_KEY or api_key in config.",
403: "Permission denied. API key may lack access to this model.",
- 404: "Model not found: '{model}'. Check spelling (e.g. 'gemini-embedding-2-preview').",
+ 404: "Model not found: '{model}'. Check spelling (e.g. 'gemini-embedding-2').",
429: "Resource exhausted. Wait and retry.",
500: "Gemini service error (Google-side). Retry after a delay.",
503: "Gemini service unavailable. Retry after a delay.",
@@ -74,7 +76,7 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
"""Dense embedder backed by Google's Gemini Embedding models.
REST endpoint: /v1beta/models/{model}:embedContent (SDK handles Parts format internally).
- Input token limit: per-model (8192 for gemini-embedding-2-preview, 2048 for gemini-embedding-001).
+ Input token limit: per-model (8192 for gemini-embedding-2, 2048 for gemini-embedding-001).
Output dimension: 1–3072 (MRL; recommended 768, 1536, 3072; default 3072).
Task types: RETRIEVAL_QUERY, RETRIEVAL_DOCUMENT, SEMANTIC_SIMILARITY, CLASSIFICATION,
CLUSTERING, CODE_RETRIEVAL_QUERY, QUESTION_ANSWERING, FACT_VERIFICATION.
@@ -82,7 +84,8 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
"""
# Default output dimensions per model (used when user does not specify `dimension`).
- # gemini-embedding-2-preview: 3072 MRL model — supports 1–3072 via output_dimensionality
+ # gemini-embedding-2: 3072 MRL model — supports 1–3072 via output_dimensionality
+ # gemini-embedding-2-preview: 3072 MRL model retained for explicit legacy configurations
# gemini-embedding-001: 3072 (native 768-dim vectors; 3072 shown as default for MRL compat)
# text-embedding-004: 768 fixed-dim legacy model, does not support MRL truncation
# Future gemini-embedding-*: default 3072 via _default_dimension() fallback
@@ -90,6 +93,7 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
supports_multimodal: bool = False # text-only; multimodal planned separately
KNOWN_DIMENSIONS: Dict[str, int] = {
+ "gemini-embedding-2": 3072,
"gemini-embedding-2-preview": 3072,
"gemini-embedding-001": 3072,
"text-embedding-004": 768,
@@ -105,8 +109,8 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
3. Fallback: 3072 (gemini-embedding-* MRL models)
Examples:
+ gemini-embedding-2 → 3072 (exact match)
gemini-embedding-2-preview → 3072 (exact match)
- gemini-embedding-2 → 3072 (fallback — future model)
text-embedding-004 → 768 (exact match)
text-embedding-005 → 768 (prefix rule — future model)
"""
@@ -118,7 +122,7 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
def __init__(
self,
- model_name: str = "gemini-embedding-2-preview",
+ model_name: str = "gemini-embedding-2",
api_key: Optional[str] = None,
dimension: Optional[int] = None,
task_type: Optional[str] = None,
@@ -149,33 +153,34 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
self._dimension = dimension or self._default_dimension(model_name)
self._token_limit = _MODEL_TOKEN_LIMITS.get(model_name, _DEFAULT_TOKEN_LIMIT)
- def _build_config(
+ def _prepare_request(
self,
+ text: str,
+ is_query: bool = False,
*,
task_type: Optional[str] = None,
title: Optional[str] = None,
- ) -> types.EmbedContentConfig:
- """Build EmbedContentConfig, merging per-call overrides with instance defaults."""
- effective_task_type = task_type or self.task_type
- kwargs: Dict[str, Any] = {"output_dimensionality": self._dimension}
- if effective_task_type:
- kwargs["task_type"] = effective_task_type.upper()
- if title:
- kwargs["title"] = title
- return types.EmbedContentConfig(**kwargs)
-
- def _resolve_task_type(
- self,
- *,
- is_query: bool = False,
- task_type: Optional[str] = None,
- ) -> Optional[str]:
+ ) -> tuple[str, types.EmbedContentConfig]:
+ """Build the model-specific text and request config."""
if task_type is None:
- if is_query and self.query_param:
- task_type = self.query_param
- elif not is_query and self.document_param:
- task_type = self.document_param
- return task_type
+ task_type = self.query_param if is_query else self.document_param
+ effective_task_type = task_type or self.task_type
+ task_type = effective_task_type.upper() if effective_task_type else None
+ config = types.EmbedContentConfig(output_dimensionality=self._dimension)
+ if self.model_name == "gemini-embedding-2":
+ if task_type and task_type not in _VALID_TASK_TYPES:
+ raise ValueError(
+ f"Invalid task_type '{effective_task_type}'. "
+ f"Valid values: {', '.join(sorted(_VALID_TASK_TYPES))}"
+ )
+ if task_type or title:
+ text = _TASK_INSTRUCTIONS[task_type or "RETRIEVAL_DOCUMENT"].format(
+ text=text, title=title or "none"
+ )
+ else:
+ config.task_type = task_type
+ config.title = title or None
+ return text, config
def __repr__(self) -> str:
return (
@@ -197,14 +202,14 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
if not text or not text.strip():
logger.warning("Empty text passed to embed(), returning zero vector")
return EmbedResult(dense_vector=[0.0] * self._dimension)
- task_type = self._resolve_task_type(is_query=is_query, task_type=task_type)
+ text, config = self._prepare_request(text, is_query, task_type=task_type, title=title)
# SDK accepts plain str; converts to REST Parts format internally.
def _call() -> EmbedResult:
result = self.client.models.embed_content(
model=self.model_name,
contents=text,
- config=self._build_config(task_type=task_type, title=title),
+ config=config,
)
vector = truncate_and_normalize(list(result.embeddings[0].values), self._dimension)
return EmbedResult(dense_vector=vector)
@@ -240,7 +245,7 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
logger.warning("Empty text passed to embed_async(), returning zero vector")
return EmbedResult(dense_vector=[0.0] * self._dimension)
- task_type = self._resolve_task_type(is_query=is_query, task_type=task_type)
+ text, config = self._prepare_request(text, is_query, task_type=task_type, title=title)
async def _call() -> EmbedResult:
# Keep creation, use and cleanup on the calling event loop.
@@ -250,7 +255,7 @@ class GeminiDenseEmbedder(DenseEmbedderBase):
result = await async_client.models.embed_content(
model=self.model_name,
contents=text,
- config=self._build_config(task_type=task_type, title=title),
+ config=config,
)
vector = truncate_and_normalize(list(result.embeddings[0].values), self._dimension)
return EmbedResult(dense_vector=vector)
diff --git a/tests/integration/conftest.py b/tests/integration/conftest.py
index bd23337d2..8d90f8a7d 100644
--- a/tests/integration/conftest.py
+++ b/tests/integration/conftest.py
@@ -52,7 +52,7 @@ requires_volcengine_kms = pytest.mark.skipif(
# (model_name, default_dimension, token_limit)
GEMINI_MODELS = [
- ("gemini-embedding-2-preview", 3072, 8192),
+ ("gemini-embedding-2", 3072, 8192),
]
@@ -84,7 +84,7 @@ def gemini_embedder():
from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
except (ImportError, ModuleNotFoundError, AttributeError):
pytest.skip("google-genai not installed")
- return GeminiDenseEmbedder("gemini-embedding-2-preview", api_key=GOOGLE_API_KEY, dimension=768)
+ return GeminiDenseEmbedder("gemini-embedding-2", api_key=GOOGLE_API_KEY, dimension=768)
def gemini_config_dict(
@@ -156,7 +156,7 @@ def sample_markdown(base_dir: Path, slug: str, content: str) -> Path:
@pytest_asyncio.fixture(scope="function")
async def gemini_ov_service(tmp_path):
- model = "gemini-embedding-2-preview"
+ model = "gemini-embedding-2"
dim = 768
service, ctx = await make_ov_service(
gemini_config_dict(model, dim),
diff --git a/tests/integration/test_gemini_e2e.py b/tests/integration/test_gemini_e2e.py
index 0b390db00..f8f0f6417 100644
--- a/tests/integration/test_gemini_e2e.py
+++ b/tests/integration/test_gemini_e2e.py
@@ -17,7 +17,7 @@ pytestmark = [pytest.mark.integration, requires_api_key]
def _cosine_similarity(a: list, b: list) -> float:
- dot = sum(x * y for x, y in zip(a, b))
+ dot = sum(x * y for x, y in zip(a, b, strict=True))
norm_a = l2_norm(a)
norm_b = l2_norm(b)
return dot / (norm_a * norm_b) if norm_a and norm_b else 0.0
@@ -26,7 +26,7 @@ def _cosine_similarity(a: list, b: list) -> float:
@pytest.fixture(scope="module")
def embedder():
e = GeminiDenseEmbedder(
- "gemini-embedding-2-preview",
+ "gemini-embedding-2",
api_key=GOOGLE_API_KEY,
# dimension defaults to 3072 — test the actual default
task_type="RETRIEVAL_DOCUMENT",
@@ -60,15 +60,17 @@ class TestGeminiE2ETextEmbedding:
sim_related = _cosine_similarity(r1.dense_vector, r2.dense_vector)
sim_unrelated = _cosine_similarity(r1.dense_vector, r3.dense_vector)
assert sim_related > sim_unrelated
+
+
class TestGeminiE2ETaskType:
def test_query_vs_document_task_types(self):
doc_embedder = GeminiDenseEmbedder(
- "gemini-embedding-2-preview",
+ "gemini-embedding-2",
api_key=GOOGLE_API_KEY,
task_type="RETRIEVAL_DOCUMENT",
)
query_embedder = GeminiDenseEmbedder(
- "gemini-embedding-2-preview",
+ "gemini-embedding-2",
api_key=GOOGLE_API_KEY,
task_type="RETRIEVAL_QUERY",
)
@@ -76,8 +78,8 @@ class TestGeminiE2ETaskType:
doc_result = doc_embedder.embed(text)
query_result = query_embedder.embed(text)
sim = _cosine_similarity(doc_result.dense_vector, query_result.dense_vector)
- # gemini-embedding-2-preview may return identical vectors for same text
- # across task types; assert vectors are at least highly correlated
+ # Stable Gemini Embedding 2 uses task-specific prompt instructions.
+ # The same source text should remain semantically correlated.
assert sim > 0.8, f"Task type similarity {sim:.3f} unexpectedly low"
doc_embedder.close()
query_embedder.close()
diff --git a/tests/integration/test_gemini_embedding_it.py b/tests/integration/test_gemini_embedding_it.py
index 00a9503d4..79a7f3cfe 100644
--- a/tests/integration/test_gemini_embedding_it.py
+++ b/tests/integration/test_gemini_embedding_it.py
@@ -56,7 +56,7 @@ def test_all_task_types_accepted(task_type):
from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
e = GeminiDenseEmbedder(
- "gemini-embedding-2-preview",
+ "gemini-embedding-2",
api_key=GOOGLE_API_KEY,
task_type=task_type,
dimension=768,
@@ -71,7 +71,7 @@ def test_config_nonsymmetric_routing():
cfg = EmbeddingConfig(
dense=EmbeddingModelConfig(
- model="gemini-embedding-2-preview",
+ model="gemini-embedding-2",
provider="gemini",
api_key=GOOGLE_API_KEY,
dimension=768,
@@ -91,7 +91,7 @@ def test_invalid_api_key_error_message():
from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
_fake_key = "INVALID_KEY_" + "XYZZY_123"
- bad = GeminiDenseEmbedder("gemini-embedding-2-preview", api_key=_fake_key)
+ bad = GeminiDenseEmbedder("gemini-embedding-2", api_key=_fake_key)
with pytest.raises(RuntimeError, match="Invalid API key"):
bad.embed("hello")
diff --git a/tests/integration/test_gemini_openviking_it.py b/tests/integration/test_gemini_openviking_it.py
index 210b86148..2ee828a50 100644
--- a/tests/integration/test_gemini_openviking_it.py
+++ b/tests/integration/test_gemini_openviking_it.py
@@ -61,7 +61,7 @@ async def test_batch_documents_search(gemini_ov_service, tmp_path):
@pytest.mark.parametrize(
"model,dim,token_limit",
[
- pytest.param("gemini-embedding-2-preview", 768, 8192, id="g2p-large"),
+ pytest.param("gemini-embedding-2", 768, 8192, id="g2-stable-large"),
pytest.param("gemini-embedding-001", 768, 2048, id="g001-large"),
],
)
@@ -99,7 +99,7 @@ async def test_large_text_add_and_search(model, dim, token_limit, tmp_path):
async def test_retrieval_routing_workflow(query_param, doc_param, tmp_path):
service, ctx = await make_ov_service(
gemini_config_dict(
- "gemini-embedding-2-preview",
+ "gemini-embedding-2",
768,
query_param=query_param,
doc_param=doc_param,
@@ -128,7 +128,7 @@ async def test_retrieval_routing_workflow(query_param, doc_param, tmp_path):
@pytest.mark.parametrize("dim", [512, 768, 1536, 3072])
async def test_dimension_variant_add_search(dim, tmp_path):
service, ctx = await make_ov_service(
- gemini_config_dict("gemini-embedding-2-preview", dim),
+ gemini_config_dict("gemini-embedding-2", dim),
str(tmp_path / f"ov_dim_{dim}"),
)
from openviking_cli.utils.config.open_viking_config import OpenVikingConfigSingleton
diff --git a/tests/unit/test_embedding_config_gemini.py b/tests/unit/test_embedding_config_gemini.py
index e1cd548ea..a8e1e7cb5 100644
--- a/tests/unit/test_embedding_config_gemini.py
+++ b/tests/unit/test_embedding_config_gemini.py
@@ -2,7 +2,7 @@
# SPDX-License-Identifier: AGPL-3.0
"""Unit tests for Gemini-specific EmbeddingModelConfig and EmbeddingConfig behavior."""
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
import pytest
@@ -12,12 +12,12 @@ from openviking_cli.utils.config.embedding_config import EmbeddingConfig, Embedd
def _gcfg(**kw) -> EmbeddingModelConfig:
"""Helper: build a Gemini EmbeddingModelConfig with defaults."""
return EmbeddingModelConfig(
- model="gemini-embedding-2-preview", provider="gemini", api_key="test-key", **kw
+ model="gemini-embedding-2", provider="gemini", api_key="test-key", **kw
)
class TestGeminiDimension:
- def test_preview_defaults_3072(self):
+ def test_stable_defaults_3072(self):
assert _gcfg().get_effective_dimension() == 3072
def test_001_defaults_3072(self):
@@ -49,14 +49,20 @@ class TestGeminiDimension:
class TestGeminiContextRouting:
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
- def test_nonsymmetric_passes_query_document_params(self, _mock):
- """get_embedder() passes query_param/document_param to GeminiDenseEmbedder."""
+ def test_nonsymmetric_routes_query_and_document(self, mock_client):
cfg = EmbeddingConfig(
dense=_gcfg(query_param="RETRIEVAL_QUERY", document_param="RETRIEVAL_DOCUMENT")
)
embedder = cfg.get_embedder()
- assert embedder.query_param == "retrieval_query"
- assert embedder.document_param == "retrieval_document"
+ request = mock_client.return_value.models.embed_content
+ request.return_value = MagicMock(embeddings=[MagicMock(values=[0.1])])
+ embedder.embed("text")
+ embedder.embed("text", is_query=True)
+ assert [call.kwargs["contents"] for call in request.call_args_list] == [
+ "title: none | text: text",
+ "task: search result | query: text",
+ ]
+ assert request.call_args.kwargs["config"].task_type is None
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
def test_only_query_param_set(self, _mock):
@@ -83,14 +89,14 @@ def test_factory_raises_clear_error_when_google_genai_not_installed():
class TestGeminiConfigValidation:
def test_missing_api_key_raises(self):
with pytest.raises(ValueError, match="api_key"):
- EmbeddingModelConfig(model="gemini-embedding-2-preview", provider="gemini")
+ EmbeddingModelConfig(model="gemini-embedding-2", provider="gemini")
def test_invalid_query_param_raises(self):
- with pytest.raises(ValueError, match="Invalid query_param"):
+ with pytest.raises(ValueError, match="invalid query_param"):
_gcfg(query_param="NOT_A_VALID_TYPE")
def test_invalid_document_param_raises(self):
- with pytest.raises(ValueError, match="Invalid document_param"):
+ with pytest.raises(ValueError, match="invalid document_param"):
_gcfg(document_param="ALSO_INVALID")
def test_query_document_param_case_normalized(self):
diff --git a/tests/unit/test_gemini_embedder.py b/tests/unit/test_gemini_embedder.py
index e5a5154fe..8fe31330c 100644
--- a/tests/unit/test_gemini_embedder.py
+++ b/tests/unit/test_gemini_embedder.py
@@ -5,7 +5,7 @@ Tests for GeminiDenseEmbedder.
Pattern: patch at module import path, use MagicMock, never make real API calls.
"""
-from unittest.mock import MagicMock, patch
+from unittest.mock import AsyncMock, MagicMock, patch
import pytest
@@ -56,7 +56,8 @@ class TestGeminiDenseEmbedderInit:
def test_default_dimension_3072(self, mock_client_class):
from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
- embedder = GeminiDenseEmbedder("gemini-embedding-2-preview", api_key="key")
+ embedder = GeminiDenseEmbedder(api_key="key")
+ assert embedder.model_name == "gemini-embedding-2"
assert embedder.get_dimension() == 3072
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
@@ -107,16 +108,17 @@ class TestGeminiDenseEmbedderEmbed:
mock_client = mock_client_class.return_value
mock_client.models.embed_content.return_value = _make_mock_result([[0.1, 0.2, 0.3]])
- embedder = GeminiDenseEmbedder("gemini-embedding-2-preview", api_key="key", dimension=3)
+ embedder = GeminiDenseEmbedder(api_key="key", dimension=3)
result = embedder.embed("hello world")
assert result.dense_vector is not None
assert len(result.dense_vector) == 3
mock_client.models.embed_content.assert_called_once()
_, kwargs = mock_client.models.embed_content.call_args
- assert kwargs["model"] == "gemini-embedding-2-preview"
+ assert kwargs["model"] == "gemini-embedding-2"
+ assert kwargs["contents"] == "hello world"
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
- def test_embed_passes_task_type_in_config(self, mock_client_class):
+ def test_preview_embed_passes_task_type_in_config(self, mock_client_class):
from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
mock_client = mock_client_class.return_value
@@ -127,9 +129,11 @@ class TestGeminiDenseEmbedderEmbed:
dimension=1,
task_type="RETRIEVAL_QUERY",
)
- embedder.embed("query text")
+ embedder.embed("query text", title="Guide")
_, kwargs = mock_client.models.embed_content.call_args
+ assert kwargs["contents"] == "query text"
assert kwargs["config"].task_type == "RETRIEVAL_QUERY"
+ assert kwargs["config"].title == "Guide"
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
def test_embed_raises_model_error_on_api_error(self, mock_client_class):
@@ -268,59 +272,32 @@ class TestGeminiErrorMessages:
embedder.embed("hello")
-class TestBuildConfig:
- @patch("openviking.models.embedder.gemini_embedders.genai.Client")
- def test_build_config_defaults(self, mock_client_class):
- from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
-
- embedder = GeminiDenseEmbedder("gemini-embedding-2-preview", api_key="key", dimension=512)
- cfg = embedder._build_config()
- assert cfg.output_dimensionality == 512
- assert cfg.task_type is None
- assert cfg.title is None
-
- @patch("openviking.models.embedder.gemini_embedders.genai.Client")
- def test_build_config_with_task_type_override(self, mock_client_class):
- from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
-
- embedder = GeminiDenseEmbedder(
- "gemini-embedding-2-preview",
- api_key="key",
- dimension=1,
- task_type="RETRIEVAL_QUERY",
- )
- cfg = embedder._build_config(task_type="SEMANTIC_SIMILARITY")
- assert cfg.task_type == "SEMANTIC_SIMILARITY"
-
- @patch("openviking.models.embedder.gemini_embedders.genai.Client")
- def test_build_config_with_title(self, mock_client_class):
- from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
-
- embedder = GeminiDenseEmbedder("gemini-embedding-2-preview", api_key="key", dimension=1)
- cfg = embedder._build_config(title="My Document")
- assert cfg.title == "My Document"
-
+class TestGeminiRequest:
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
def test_embed_per_call_task_type(self, mock_client_class):
from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
mock_client = mock_client_class.return_value
mock_client.models.embed_content.return_value = _make_mock_result([[0.1]])
- embedder = GeminiDenseEmbedder("gemini-embedding-2-preview", api_key="key", dimension=1)
- embedder.embed("text", task_type="CLUSTERING")
+ embedder = GeminiDenseEmbedder(api_key="key", dimension=1, task_type="RETRIEVAL_QUERY")
+ embedder.embed("text", task_type="clustering")
_, kwargs = mock_client.models.embed_content.call_args
- assert kwargs["config"].task_type == "CLUSTERING"
+ assert kwargs["contents"] == "task: clustering | query: text"
+ assert kwargs["config"].model_dump(exclude_none=True) == {"output_dimensionality": 1}
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
- def test_embed_per_call_title(self, mock_client_class):
+ async def test_embed_async_per_call_title(self, mock_client_class):
from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
- mock_client = mock_client_class.return_value
- mock_client.models.embed_content.return_value = _make_mock_result([[0.1]])
- embedder = GeminiDenseEmbedder("gemini-embedding-2-preview", api_key="key", dimension=1)
- embedder.embed("text", title="Doc Title")
- _, kwargs = mock_client.models.embed_content.call_args
- assert kwargs["config"].title == "Doc Title"
+ mock_client = mock_client_class.return_value.__enter__.return_value
+ mock_async = mock_client.aio.__aenter__.return_value
+ mock_async.models.embed_content = AsyncMock(return_value=_make_mock_result([[0.1]]))
+ embedder = GeminiDenseEmbedder(api_key="key", dimension=1)
+ result = await embedder.embed_async("text", title="Doc Title")
+ kwargs = mock_async.models.embed_content.call_args.kwargs
+ assert result.dense_vector == [0.1]
+ assert kwargs["contents"] == "title: Doc Title | text: text"
+ assert kwargs["config"].model_dump(exclude_none=True) == {"output_dimensionality": 1}
@patch("openviking.models.embedder.gemini_embedders.genai.Client")
def test_repr(self, mock_client_class):
@@ -345,7 +322,7 @@ class TestBuildConfig:
GeminiDenseEmbedder,
)
- GeminiDenseEmbedder("gemini-embedding-2-preview", api_key="key")
+ GeminiDenseEmbedder(api_key="key")
mock_client_class.assert_called_once()
call_kwargs = mock_client_class.call_args[1]
assert call_kwargs.get("api_key") == "key"