Question : API

[theapache64]

I've been developing a telegram bot using the API. Currently am using https://covid-middleware.deepset.ai/api/bert/question to get the answers.

curl -X POST \
  https://covid-middleware.deepset.ai/api/bert/question \
  -H 'content-type: application/json' \
  -d '{
    "question":"community spread?"
}'

but the swagger doesn't list this API and shows a different one with different structures.

So, my question is, Which API should I choose to get the answers? @tanaysoni

[tanaysoni]

Hi @theapache64, during the hackathon, the middleware was deprecated. The new updated APIs are available at https://covid-backend.deepset.ai. The documentation is at https://covid-backend.deepset.ai/docs.

I hope that clears the confusion.

[theapache64]

Alright.

Can you explain the parameters ?

{
  "questions": [
    "string"
  ],
  "filters": {
    "additionalProp1": "string",
    "additionalProp2": "string",
    "additionalProp3": "string"
  },
  "top_k_reader": 0,
  "top_k_retriever": 0
}

[sfakir]

@tanaysoni I propose we hide the "model" from the the clients. because it's not responsibility of the clients to know what kind of models were used in the backend.

I create an MR to have a unified endpoint and internally we route to the right model, what do you think?

[tanaysoni]

@theapache64, a sample request looks like this:

{
    "questions": ["what are the symptoms?"],
    "top_k_retriever": 5 // return top 5 answers
}

The top_k_reader parameter can be ignored for FAQ style Question Answering.

With the current version, filters are tricky to implement as there's no endpoint to get a list of available filters. Feel free to create a PR if that's something useful for your use case!

[sfakir]

@tanaysoni In MR74 I added POST /question/ask as a proxy endpoint for our models.

It's my recommendation to have this very abstract endpoint we can keep and there won't be a change later, also we are able then to change the model internally without notifying all clients. We can have a chat about this later on.

[theapache64]

@tanaysoni Which model should I use ? 1 or 2 ?

[tanaysoni]

@sfakir a unified endpoint sounds good to me!

I would propose to keep both, the unified endpoint and model-specific endpoints.

In the unified endpoint, we detect language of the request and route it to the appropriate model. For advanced use cases, clients can choose between the model-specific endpoints.

[tanaysoni]

@theapache64 model 1 is multi-lingual, while model 2 is exclusive English. If you're working only with English then model 2 might have better results.

[theapache64]

@tanaysoni Which languages are supported in model 1?

[tholor]

In theory, it's multilingual. In practice, we only have reasonable amount of FAQ data for German and English right now.

We are currently switching to a generic API endpoint (see https://github.com/deepset-ai/COVID-QA/pull/72). So please use /question/ask once it's deployed (without any model id). We will then take care of routing it to the best model available for the detected language.