hayhooks: new Pipeline deployment

[mpangrazzi]

Motivation

Currently, the /deploy endpoint is accepting pipeline's YAML (as string) and pipeline's name. This seems ok, but as we know it also causes many issues while trying to create Pydantic models for component's inputs/outputs.

If we want a smooth integration with open-webui, we need to consider that the user's input will be basically a message (in OpenAI format) and nothing else.

Currently, if we deploy this pipeline, to call it we need to provide this JSON body:

{
  "converter": {
    "meta": {},
    "extraction_kwargs": {}
  },
  "fetcher": {
    "urls": [
      "https://www.redis.io"
    ]
  },
  "llm": {
    "system_prompt": "string",
    "generation_kwargs": {}
  },
  "prompt": {
    "query": "What is redis?",
    "template": "string",
    "template_variables": {}
  }
}

So the problem is: this pipeline accept multiple inputs (e.g. prompt.query and/or fetcher.urls) while we have one message coming from the user.

To solve this, I would implement first a deployment approach similar to the one we saw in Pipelines, where the pipeline execution is defined on a user's defined method of a wrapper class, having the user's message as an input.