ml-commons
ml-commons copied to clipboard
opensearch-project
[Enhancement] Simplify the current user experience and APIs in Ml-Commons
From customer feedbacks, building ML apps on top of Ml-Commons requires too many upfront setup or pre-requisite APIs calls that looks unnecessary or unwanted to them. Some improvement are done in https://github.com/opensearch-project/ml-commons/issues/1147 and https://github.com/opensearch-project/ml-commons/issues/1146 recently. However, there are still more to enhance to lower the bar for both internal and external customers to onboard Ml-Commons.
We need to propose a roadmap to simplify as much as possible to help customers, starting with the below pain points:
- Allow remote inference without registering models. Users just want to create connector and assume it in their applications to run remote "Predict". - this is achieved by flow framework
- Remove the obstacle that are brought in by model_groups. - this is resolved by default model_group
- Get ride of manually deploy/undeploy models that are required in the workflow. https://github.com/opensearch-project/ml-commons/issues/1148 - this is resolved in 2.14 release.
Feel free to add more thoughts and suggestions!
One more point: add blueprint id, so user can just input "cohere_embedding" , rather than copy the whole blue print
This is the exact case we are working on. To simplify the setup for registering the model. Please take a look of https://github.com/opensearch-project/opensearch-ai-flow-framework/issues/21 and https://github.com/opensearch-project/opensearch-ai-flow-framework/issues/70. We are trying to automate creating a connector, model group then registering and deploying the remote model.
@owaiskazi19 as part of the AI Flow project, are you also planning to simplify the APIs as well? Or, are you just creating a UI that interacts with the existing APIs? Thanks!
@owaiskazi19 as part of the AI Flow project, are you also planning to simplify the APIs as well? Or, are you just creating a UI that interacts with the existing APIs? Thanks!
@jonfritz we are creating wrapper APIs which would simplify the provision of complex setups of OpenSearch. Details for the API can be found here. AI Flow project will have a separate backend and not just frontend. User can use the APIs to create a template based on the use case and execute it.
@Zhangxunmt please see https://github.com/opensearch-project/ml-commons/issues/1343#issuecomment-1753428715 by @owaiskazi19 to address the problems you highlighted in this issue.
