Feature Request: support to add custom classification categories

[vs9205]

Is your feature request related to a problem? Please describe.

Currently, the classification model modernBERT is pre-trained to classify input prompts into a fixed set of 9 categories such as business, law, history, etc. This limits flexibility when users want to route inputs based on other task-specific categories like summarization, translation, named entity recognition (NER), sentiment analysis, and more.

Describe the solution you'd like

To enhance flexibility, we propose enabling support for custom classification categories by allowing users to fine-tune the modernBERT model on their own datasets. This could be achieved by extending the configuration file to accept a URL (e.g., a Hugging Face dataset link or a custom dataset) for fine-tuning modernBERT. The system would then use this fine-tuned model instead of the default pre-trained model for classification tasks. This approach empowers users to tailor the semantic routing to their specific domain and task requirements without being constrained to the predefined categories.

[rootfs]

For the model training and release process, ref #193

[QuanHNguyen232]

@rootfs This issue is interesting. I'd love to handle it. Can you assign it to me? I'm drafting a plan for it and will update with you.

[rootfs]

@QuanHNguyen232 thank you for your interests! we have received many interests on custom filter. I'll create a high level design doc and create sub tasks.

[rootfs]

@QuanHNguyen232 please review #358