[Prompt Classification] Implement In-Tree Embedding Similarity Matching

[rootfs]

Goal

Implement prompt classification using embedding similarity.

Acceptance Criteria

• Reuse existing candle-binding embedding models
• Embedding similarity scoring
• Configurable thresholds per category (e.g. category A: 80%, category B: 60%)
• Support mean/max/any aggregation methods to find the best match
• Unit tests results
• Benchmark results

[Sophie8]

@rootfs Hi this feature I am interested in having a try, will provide a write up soon and posted here. also saw the example embedding server for the out of tree mcp here: https://github.com/vllm-project/semantic-router/pull/417/files#diff-1f55fea4489767724382d45c8e90a41f85d05ffe166e742f381cfad49df68e1c, do we expect to centralize all the embedding models later?

[rootfs]

@Sophie8 thanks! yes, that is the flow that can go to the in-tree classification. We are trying to support the same embedding models (qwen3 embedding and gemma) in #266 . You can use feat-candle-refactoring for development first and send the PR to that branch.

[Sophie8]

@Sophie8 thanks! yes, that is the flow that can go to the in-tree classification. We are trying to support the same embedding models (qwen3 embedding and gemma) in #266 . You can use feat-candle-refactoring for development first and send the PR to that branch.

Thank you! let me check out that branch and get startted!