[Feature][Integration][Java] Add Milvus vector store integration

[yanbinyang]

Linked issue: #661

Purpose of change

This PR implements the Java version of Milvus-based Vector Store functionality for Flink Agents, following the design proposal in #143. This implementation enables RAG (Retrieval-Augmented Generation) capabilities by providing vector-based context retrieval.

• Added Milvus vector store integration
• Added Milvus vector store module under integrations/vector-stores
• Supported Milvus connection and query configuration
• Converted Milvus search results into the common Document format
• Added tests for Milvus vector store behavior

Tests

API

Documentation

• doc-needed
• doc-not-needed
• doc-included

[wenjin272]

Thanks for your contribution @yanbinyang. Overall looks good to me, and I left serveral minor comments.

I think we should also verify Milvus in the test case Mem0LongTermMemoryTest. This test is currently disabled due to a bug in Flink itself, which is expected to be fixed in the coming weeks. There are also workarounds available for local execution. You can add the test path for Milvus, and I will verify it locally.

[wenjin272]

It looks like there were some conflicts during the rebase. In a previous PR, we removed the Collection data structure and this line of description.

[yanbinyang]

It looks like there were some conflicts during the rebase. In a previous PR, we removed the Collection data structure and this line of description.

Thanks for catching this. Removed the stale Collection concept line in ae1261a

[wenjin272]

Since flink-agents supports asynchronous execution of actions, meaning that multiple threads can invoke the same vector store simultaneously, connection leaks caused by multi-threading may occur here.

I suggest adding DCL.

[yanbinyang]

Since flink-agents supports asynchronous execution of actions, meaning that multiple threads can invoke the same vector store simultaneously, connection leaks caused by multi-threading may occur here.

I suggest adding DCL.

Good point. Addressed in 56cc845 by making the Milvus client volatile and using double-checked locking for lazy initialization. I also synchronized close() so concurrent action execution will not create or leak multiple clients.

[wenjin272]

It appears that this file is not being used in the CI. I'm wondering if we could also align the current Elasticsearch startup process in the CI with that of Milvus.

[yanbinyang]

It appears that this file is not being used in the CI. I'm wondering if we could also align the current Elasticsearch startup process in the CI with that of Milvus.

Addressed in 37c5436. The cross-language CI job now starts and stops Elasticsearch via tools/docker/elasticsearch/docker-compose.yml, aligned with the Milvus startup flow.

[yanbinyang]

Thanks for your contribution @yanbinyang. Overall looks good to me, and I left serveral minor comments.

I think we should also verify Milvus in the test case Mem0LongTermMemoryTest. This test is currently disabled due to a bug in Flink itself, which is expected to be fixed in the coming weeks. There are also workarounds available for local execution. You can add the test path for Milvus, and I will verify it locally.

Thanks for the suggestion. Addressed in fa34054 by adding a Milvus path to Mem0LongTermMemoryTest: the test now parameterizes the Mem0 vector store over Elasticsearch and Milvus, and Mem0LongTermMemoryAgent declares a milvusLtmStore resource.

I couldn't run this e2e test locally because it still depends on the existing Pemja/Flink issue workaround and a local Milvus/Mem0 environment. Could you please help verify the Milvus path locally?