Skip to content

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

3.1 KB
Raw Blame
# Copyright (c) 2026 Beijing Volcano Engine Technology Co., Ltd.
# SPDX-License-Identifier: AGPL-3.0
"""
End-to-end integration tests for GeminiDenseEmbedder.
Calls the real Gemini API — requires GOOGLE_API_KEY env var.
Run: pytest tests/integration/test_gemini_e2e.py -v -m integration
"""

import pytest

pytest.importorskip("google.genai")

from openviking.models.embedder.gemini_embedders import GeminiDenseEmbedder
from tests.integration.conftest import GOOGLE_API_KEY, l2_norm, requires_api_key

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, 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


@pytest.fixture(scope="module")
def embedder():
    e = GeminiDenseEmbedder(
        "gemini-embedding-2",
        api_key=GOOGLE_API_KEY,
        # dimension defaults to 3072 — test the actual default
        task_type="RETRIEVAL_DOCUMENT",
    )
    yield e
    e.close()


def test_default_dimension_is_3072(embedder):
    """Default output dimension must match model's native 3072."""
    assert embedder.get_dimension() == 3072
    result = embedder.embed("hello")
    assert len(result.dense_vector) == 3072


class TestGeminiE2ETextEmbedding:
    def test_embed_text_returns_correct_dimension(self, embedder):
        result = embedder.embed("OpenViking is a knowledge management system")
        assert result.dense_vector is not None
        assert len(result.dense_vector) == 3072

    def test_embed_text_vector_is_normalized(self, embedder):
        result = embedder.embed("test normalization")
        norm = l2_norm(result.dense_vector)
        assert abs(norm - 1.0) < 0.01, f"Vector norm {norm} not close to 1.0"

    def test_semantic_similarity_related_texts(self, embedder):
        r1 = embedder.embed("a golden retriever playing in the park")
        r2 = embedder.embed("a dog running outside in a field")
        r3 = embedder.embed("quantum computing and cryptography")
        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",
            api_key=GOOGLE_API_KEY,
            task_type="RETRIEVAL_DOCUMENT",
        )
        query_embedder = GeminiDenseEmbedder(
            "gemini-embedding-2",
            api_key=GOOGLE_API_KEY,
            task_type="RETRIEVAL_QUERY",
        )
        text = "machine learning algorithms"
        doc_result = doc_embedder.embed(text)
        query_result = query_embedder.embed(text)
        sim = _cosine_similarity(doc_result.dense_vector, query_result.dense_vector)
        # 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()