Some prompts may merit a recipe that is a set of horizontal steps, ie a prompt such as "I want compounds with activity against malaria" could return a number of models from the hub, each of which can be run independently of each other. These models do not require to be run in any particular order, hence we call this kind of recipe "horizontal".
In contrast to this, we define "vertical" or sequential recipes. For example, for the following prompt "I want compounds with activity against malaria with higher pk profile", the recipe should involve first filtering out all compounds that are active against malaria, and then determining their pk profile.
Generally, given the fact that the hub has more than one model for most pathogens and ADMET tasks, we expect most queries to be a combination of sequential task where each step might involve more than one relevant model.
We want to carry out two things here:
Identify to which category a user prompt belongs.
For each task in the user prompt, return a cluster of relevant models.
Some prompts may merit a recipe that is a set of horizontal steps, ie a prompt such as "I want compounds with activity against malaria" could return a number of models from the hub, each of which can be run independently of each other. These models do not require to be run in any particular order, hence we call this kind of recipe "horizontal".
In contrast to this, we define "vertical" or sequential recipes. For example, for the following prompt "I want compounds with activity against malaria with higher pk profile", the recipe should involve first filtering out all compounds that are active against malaria, and then determining their pk profile.
Generally, given the fact that the hub has more than one model for most pathogens and ADMET tasks, we expect most queries to be a combination of sequential task where each step might involve more than one relevant model.
We want to carry out two things here: