Isolate Milvus collections by embedding model

[LiemFrans]

Description

Prevent Milvus startup failures when an existing unsuffixed collection was created with a different embedding model or vector dimension. This aligns Milvus with the model-aware isolation already used by other vector backends while still preserving compatible legacy collections.

Related Issues

• No tracked issue number.
• Addresses the Milvus startup failure caused by Vector dimension mismatch for collection 'entities': existing=4096, current=2560.

Changes Made

• Added model-aware Milvus collection naming based on embedding model and dimension.
• Reused compatible legacy Milvus collections when an existing unsuffixed collection still matches the current schema and embedding dimension.
• Created a new suffixed collection when a legacy collection exists but is incompatible with the current embedding configuration.
• Hardened collection suffix generation to ignore non-string model_name values.
• Added regression tests covering suffixed naming, compatible legacy reuse, and incompatible legacy fallback behavior.

Checklist

• Changes tested locally
• Code reviewed
• Documentation updated (if necessary)
• Unit tests added (if applicable)

Additional Notes

• Validated with ./scripts/test.sh tests/test_milvus_index_creation.py.
• Focused test run passed: 17 tests.
• PR target should be HKUDS/LightRAG:main.
• PR head should be LiemFrans:fix/milvus-model-isolation.

[danielaskdd]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7aaf56c436

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Skip legacy reuse when embedding model identity is unknown

This branch reuses the unsuffixed legacy collection whenever _activate_existing_collection succeeds, but that compatibility path validates only schema and vector dimension, not the embedding model identity. In a migration from one model to another with the same dimension (for example, two different 1536d models), initialization will silently bind to the old legacy collection and mix incompatible vectors, which undermines retrieval correctness and defeats the model-isolation goal.

Useful? React with 👍 / 👎.