feat: Add Milvus vector store backend support for RAG Engine

[bangqipropel]

• Add MilvusVectorStoreHandler with full CRUD overrides (list, update, delete, dedup, retrieve)
• Wire Milvus routing in main.py and config.py
• Add pymilvus and llama-index-vector-stores-milvus dependencies
• Add unit tests (test_milvus_store.py) using Milvus Lite
• Add Milvus deployment YAML and RAGEngine example YAML
• Add Milvus E2E test case and workflow deployment step
• Update CRD comments, docs (rag.md, rag-api.md) with Milvus info

Reason for Change:

Requirements

• added unit tests and e2e tests (if applicable).

Issue Fixed:

Notes for Reviewers:

[kaito-pr-agent]

Title

Add Milvus vector store backend support for RAG Engine

---

Description

• Implement MilvusVectorStoreHandler with full CRUD and index restoration logic.

• Integrate Milvus routing in main entry point and configuration files.

• Add unit tests and E2E test cases for Milvus backend validation.

• Update documentation and provide deployment example manifests.

---

Changes walkthrough 📝

	Relevant files
Configuration changes	3 files ragengine_types.go Update CRD comments to include Milvus support                        +3/-2      config.py Update supported vector DB types in comments                          +1/-1      ragengine-e2e-workflow.yaml Add Milvus deployment step in CI                                                  +51/-0
Tests	2 files rag_test.go Add E2E test suite for Milvus backend                                        +140/-0  test_milvus_store.py Add unit tests for Milvus store                                                    +170/-0
Enhancement	2 files main.py Add Milvus handler initialization logic                                    +9/-1      milvus_store.py Implement Milvus vector store handler                                        +690/-0
Documentation	4 files kaito_ragengine_milvus.yaml Add RAGEngine example for Milvus                                                  +19/-0    milvus-deployment.yaml Add Milvus deployment manifest                                                      +212/-0  rag-api.md Update API comparison table with Milvus                                    +8/-8      rag.md Add Milvus setup and usage documentation                                  +53/-0
Dependencies	1 files requirements.txt Add Milvus Python dependencies                                                      +2/-0

---

Need help?

Type /help how to ... in the comments thread for any questions about PR-Agent usage.

Check out the documentation for more information.

[kaito-pr-agent]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 PR contains tests

🔒 Security concerns Sensitive information exposure: The milvus-deployment.yaml file contains hardcoded MinIO credentials (minioadmin/minioadmin) that should be moved to Kubernetes Secrets. Additionally, the VECTOR_DB_ACCESS_SECRET is referenced but the actual secret creation is not shown in the deployment examples.

⚡ Recommended focus areas for review Critical Bug - Wrong Backend in Test The first Milvus test case (lines 420-475) incorrectly calls createLocalEmbeddingKaitoVLLMRAGEngineWithQdrant instead of createLocalEmbeddingKaitoVLLMRAGEngineWithMilvus. This is a copy-paste error that would test Qdrant instead of Milvus, defeating the purpose of the test. ragengineObj := createLocalEmbeddingKaitoVLLMRAGEngineWithQdrant(clusterIP, "v1/completions") Performance Concern - Pagination Efficiency The _list_documents_in_index method over-fetches entities (batch_size = max(limit * 3, 100)) and deduplicates in Python. For large collections with many chunks per document, this could result in excessive Milvus queries and memory usage. Consider implementing server-side deduplication or more efficient pagination. Security - Hardcoded Credentials The MinIO deployment uses hardcoded credentials (MINIO_ACCESS_KEY: minioadmin, MINIO_SECRET_KEY: minioadmin) in the YAML. These should be stored in Kubernetes Secrets and referenced via envFrom or secretKeyRef to prevent credential exposure in version control. - name: MINIO_ACCESS_KEY value: minioadmin - name: MINIO_SECRET_KEY value: minioadmin

[kaito-pr-agent]

PR Code Suggestions ✨

Explore these optional code suggestions:

Category	Suggestion	Impact
Possible issue	Fix incorrect backend function call in Milvus test case This test case is named "should create RAG with Milvus vector database backend successfully" but incorrectly calls the Qdrant factory function. Replace it with the Milvus-specific function to test the correct backend. test/rage2e/rag_test.go [442] -ragengineObj := createLocalEmbeddingKaitoVLLMRAGEngineWithQdrant(clusterIP, "v1/completions") +ragengineObj := createLocalEmbeddingKaitoVLLMRAGEngineWithMilvus(clusterIP, "v1/completions") Suggestion importance[1-10]: 9 __ Why: The test case is named "should create RAG with Milvus vector database backend successfully" but uses the Qdrant factory function. This is a critical bug where the test code does not match the test intent/name.	High
	Fix hash comparison to use consistent metadata The hash comparison uses empty metadata for the old document but includes metadata for the new document. This inconsistency may cause false positives where unchanged documents are marked as updated. Use consistent metadata for both hash calculations. presets/ragengine/vector_store/milvus_store.py [546-552] # Compare hash to skip unchanged documents old_text = result[0].get("text", "") +old_metadata = result[0].get("metadata", {}) new_hash = LlamaDocument(text=doc.text, metadata=doc.metadata).hash -old_hash = LlamaDocument(text=old_text, metadata={}).hash +old_hash = LlamaDocument(text=old_text, metadata=old_metadata).hash if new_hash == old_hash: unchanged_docs.append(doc) continue Suggestion importance[1-10]: 2 __ Why: While the suggestion correctly identifies a metadata inconsistency in hash comparison, the proposed fix is ineffective because the query in update_documents only requests ["text"], so metadata will not be present in result[0] regardless of the code change.	Low
General	Correct nodeSelector workload label for Milvus components The nodeSelector uses workload: qdrant for Milvus components (etcd, minio, milvus), but this should be workload: milvus to match the intended workload label and avoid scheduling conflicts with Qdrant deployments. examples/RAG/milvus-deployment.yaml [28-29] nodeSelector: - workload: qdrant + workload: milvus Suggestion importance[1-10]: 9 __ Why: Copy-paste error from Qdrant manifest causes incorrect node scheduling labels, likely preventing deployment if nodes are not labeled qdrant.	High
	Update retrieval comment to reflect Milvus search behavior The comment references Qdrant's hybrid search feature, but this test is for Milvus which uses dense vector similarity. Update the comment to accurately reflect Milvus's search behavior. test/rage2e/rag_test.go [459-461] -// Retrieve document - Qdrant uses hybrid search (dense + sparse) +// Retrieve document - Milvus uses dense vector similarity expectedText := indexDoc["text"].(string) err = createAndValidateRetrievalPod(ragengineObj, docID, expectedText) Suggestion importance[1-10]: 7 __ Why: The comment references Qdrant's hybrid search, but the test is for Milvus. Updating the comment improves accuracy and consistency with the test name.	Medium
	Align E2E Milvus deployment with production configuration The E2E Milvus deployment uses embedded etcd and local storage, which differs from the production manifest that uses external etcd and MinIO. This inconsistency may cause test failures that don't reflect production behavior. Consider aligning the E2E setup with production or documenting the difference. .github/workflows/ragengine-e2e-workflow.yaml [195-199] env: -- name: ETCD_USE_EMBED - value: "true" -- name: COMMON_STORAGETYPE - value: "local" +- name: ETCD_ENDPOINTS + value: etcd:2379 +- name: MINIO_ADDRESS + value: minio:9000 Suggestion importance[1-10]: 5 __ Why: Points out configuration inconsistency between E2E and production manifests, which is valid but E2E often prioritizes simplicity.	Low
	Update Milvus URL to use correct namespace The documentation shows Milvus URL using default namespace, but the E2E workflow deploys Milvus to the milvus namespace. Update the documentation to reflect the correct namespace or clarify that users should adjust the URL based on their deployment namespace. website/docs/rag.md [229] -url: "http://milvus.default.svc.cluster.local:19530" +url: "http://milvus.milvus.svc.cluster.local:19530" Suggestion importance[1-10]: 5 __ Why: Improves consistency between documentation and E2E test configuration, though default namespace is also acceptable for examples.	Low
Security	Replace hardcoded MinIO credentials with Secret references The MinIO deployment uses default credentials (minioadmin/minioadmin) which is a security risk. Replace these with environment variable references to Kubernetes Secrets to avoid exposing credentials in the manifest. examples/RAG/milvus-deployment.yaml [102-106] env: - name: MINIO_ACCESS_KEY - value: minioadmin + valueFrom: + secretKeyRef: + name: minio-credentials + key: access-key - name: MINIO_SECRET_KEY - value: minioadmin + valueFrom: + secretKeyRef: + name: minio-credentials + key: secret-key Suggestion importance[1-10]: 8 __ Why: Hardcoded credentials in example manifests are a security risk. Using Kubernetes Secrets is best practice even for examples.	Medium

[bfoley13]

is this and minio needed here? if so might not want the qdrant nodeSelector

[bfoley13]

im curious what kind of latency this will add on startup? we might want to have some telemetry on this

[bfoley13]

is the total count actually an approximate? seems like count(*) will take care of this?

[bfoley13]

looks like this is pointing to qdrant

[bfoley13]

Ahh I see that the etcd + minio is used by milvus in the deployment example, so just remove the nodeSelectors from the example deployment