Updated the integration coverage with live embedding-provider E2E tests for both Milvus Lite and Zilliz Cloud, while keeping the deterministic embedding tests for stable local/CI smoke coverage.

Local validation:

• uv run pytest tests/unittests/memory/test_milvus_memory_service.py -> 12 passed
• RUN_MILVUS_LITE_E2E=1 uv run pytest tests/integration/memory/test_milvus_memory_service_e2e.py -q -> 4 passed, 2 skipped
• uv run pytest -> 88 passed, 18 skipped

The skipped live embedding checks are for a provider endpoint that is unavailable from the current location.

Tightened the live integration tests so external services are fully opt-in. Having credentials in the environment is no longer enough to trigger network-backed Milvus/Zilliz/OpenAI/Google checks; callers must also set the relevant RUN_*_E2E=1 flags.

I also expanded the Milvus memory sample to clarify that the service accepts any embedding function, with both Gemini and OpenAI embedding examples.

Local validation:

• uv run pytest tests/integration/memory/test_milvus_memory_service_e2e.py -q -> 6 skipped
• RUN_MILVUS_LITE_E2E=1 RUN_ZILLIZ_CLOUD_E2E=1 RUN_OPENAI_EMBEDDING_E2E=1 RUN_GOOGLE_EMBEDDING_E2E=1 uv run pytest tests/integration/memory/test_milvus_memory_service_e2e.py -q -> 4 passed, 2 skipped
• uv run pytest tests/unittests/memory/test_milvus_memory_service.py -> 12 passed
• uv run pytest -> 86 passed, 20 skipped

Added the RAG side of the Milvus integration in a separate commit.

What changed:

• Added MilvusVectorStore for collection setup, schema validation, text ingestion, sync/async similarity search, and configurable URI/token/db/consistency settings.
• Added MilvusToolset with a prefixed milvus_similarity_search retrieval tool.
• Added unit coverage plus opt-in live E2E coverage for Milvus Lite and Zilliz Cloud.
• Added a sample README for using the toolset with ADK agents.

Validation:

• uv run --with isort isort src/google/adk_community/tools/milvus tests/unittests/tools/milvus tests/integration/tools/test_milvus_toolset_e2e.py
• uv run --with pyink pyink --config pyproject.toml src/google/adk_community/tools/milvus tests/unittests/tools/milvus tests/integration/tools/test_milvus_toolset_e2e.py
• uv run pytest tests/unittests/tools/milvus/test_milvus_toolset.py -> 11 passed
• uv run pytest tests/integration/tools/test_milvus_toolset_e2e.py -q -> 6 skipped by default
• Opt-in live E2E against Milvus Lite and Zilliz Cloud with live embedding checks -> 4 passed, 2 skipped. The skipped cases are the Google embedding checks because that API is not available from the current test location.
• uv run pytest tests/unittests/tools -> 32 passed
• uv run pytest -> 97 passed, 26 skipped

Hi @DeanChensj and @secprog, this is ready for review when you have a chance.

This implements the Milvus memory and RAG toolset support in the community repo, including the use case from #70, with opt-in live E2E coverage for Milvus Lite and Zilliz Cloud. The companion docs PR is google/adk-docs#1924.

Hi @DeanChensj @secprog, friendly follow-up when you have a chance.

The PR now includes both Milvus memory service support and the Milvus RAG toolset, with deterministic Milvus Lite coverage plus opt-in live Milvus Lite / Zilliz Cloud embedding-provider E2E checks. Current checks are passing.

Closing this PR because the Milvus ADK integration is now maintained as the standalone adk-milvus package:

• GitHub: https://github.com/zilliztech/adk-milvus
• PyPI: https://pypi.org/project/adk-milvus/

The companion docs PR has been updated to point to the standalone package instead of this community-repo implementation. This avoids maintaining two parallel integration surfaces. Happy to revisit a smaller wrapper or adapter here later if the community repo wants an in-repo entry point.