[Feedback] Machine Learning Model Serving Framework - Experimental Release

Good to know that OpenSearch will be more extensible and integrated with external systems. This will really help a variety of usecases. Regarding the RAG, is this going to be built-in in ml-commons or will it be supported through external APIs?

Yes, the ML extensibility capabilities are part of ml-commons. The external APIs will be mapped to the ml-commons APIs. The models managed by ml-commons will be an abstraction that could be associated with local and external models. The RAG workflow will be achieved through a workflow configuration that allows you to select the source of your embedding and generative models. You could use a self-managed generative model that you trained on your own or you could leverage a foundation model service.

Having an issue where i uploaded a model and tested the neural search plugin and it worked great. Then i deleted the model using the API’s but i keep seeing these errors in the log

ML model not in cache. Remove all of its cache files.

Cant tell what is looking for the deleted model ? I even deleted all the indexes and pipelines i had created. Any insights on what could be happening or what to look for ?

@srikant This log means there are some local cache folder not deleted yet for the model, the cron job trying to delete the file cache but seems these cache files can’t be deleted.

Which version are you using? You can check your <OpenSearch_home_folder>/data folder, you can manually delete the ml_cache folder, or djl folder if you are using version before 2.7.

Great, thanks that worked.

Running into an issue where I’m trying to store a vector size of 1536, and I’m getting an error “failed to parse”. The parser gets to a certain part of the vector and just stops. How can I solve this?

hi, @rid, can you share more details like what’s the index setting? You can also test if the model can return correct embedding by calling predict API , refer to Model-serving framework - OpenSearch documentation