MLflow model adapter

[fedefernandez]

Overview

This PR adds a new Ktor client's plugin for adapting the Xef request/response model (OpenAI compatible models) to MLflow.

sequenceDiagram
    actor Client
    participant A as XefServer
    participant B as MLflowModelAdapter
    participant C as MLflowGateway
    Client->>+A: OpenAI Request
    A->>+B: OpenAI Request
    B->>+C: MLflow Request
    C-->>-B: MLflow Response
    B-->>-A: OpenAI Response
    A-->>-Client: OpenAI Response

Change details

1. Simplify ModelUriAdapter.kt by intercepting the "request pipeline" (instead of "send pipeline") since we only need to change the request
2. Add the models that will be used for transforming Xef requests into MLflow requests and MLflow responses into Xef responses.
3. Add the MLflowModelAdapter.kt and the builder

How to test it?

1. Follow the Quickstart Guide for starting an MLflow Gateway instance
2. Add the interceptors to the Ktor HTTP client in Server

install(com.xebia.functional.xef.server.http.client.ModelUriAdapter) {
  addToPath(
    com.xebia.functional.xef.server.http.client.OpenAIPathType.EMBEDDINGS, 
    "embedding-model" to "http://127.0.0.1:5000/gateway/embeddings/invocations"
  )
  addToPath(
    com.xebia.functional.xef.server.http.client.OpenAIPathType.CHAT, 
    "chat-model" to "http://127.0.0.1:5000/gateway/chat/invocations"
  )
}
install(com.xebia.functional.xef.server.http.client.mlflow.MLflowModelAdapter) {
  addToPath(
    "http://127.0.0.1:5000/gateway/embeddings/invocations",
    com.xebia.functional.xef.server.http.client.OpenAIPathType.EMBEDDINGS
  )
  addToPath(
    "http://127.0.0.1:5000/gateway/chat/invocations",
    com.xebia.functional.xef.server.http.client.OpenAIPathType.CHAT
  )
}

Next Steps

Necessary code for automatically starting Xef server with an MLflow Gateway connection through configuration and presumably just with an environment variable.

[raulraja]

Looks great @fedefernandez !