Introduction to Fluidity, a Python class library for easy LLM workflow development

[ash.thakur]

I would like to introduce Fluidity (GitHub: AshT-Python/Fluidity & pip install fluidityai), a set of Python classes that makes AI workflow development particularly easy and intuitive.

It’s best to illustrate with an example:

w = Workflow(“Monarchs”)
t1 = Task(name=“King”, instructions=“Please give me the name of a famous king”, output_type=“string”)
t2 = Task(name=“Queen”, instructions=“Please give me the name of a famous queen”, output_type=“string”)
t3 = Task(name=“DOB”, instructions=“Please tell me the dates of birth of {King} and {Queen} in the format dd-Mmm-yyyy”, output_type=“dict”)
t4 = Task(name=“Older”, instructions=“Please tell me which of these two monarchs was older, given their dates of birth: {DOB}”)
t1.setNextTask(“Queen”)
t2.setNextTask(“DOB”)
t3.setNextTask(“Older”)
w.addTasks(t1, t2, t3, t4)
w.run(“King”) # Run workflow from specified task name
print(w.outputs()[“Older”]) # prints the name of the older monarch

Essentially you create a workflow and add tasks to it. Each task points to its next task - alternatively, branch objects can be added to the workflow to programmatically decide where to go next by setting the “next_task” pointer. This way potentially complex workflow graphs, which can include cycles, can be constructed easily. The output of any task can be accessed by subsequent tasks using {<task_name>} placeholders. A task can be python code instead of LLM instructions.

[ash.thakur]

Btw there are lots of useful subclasses of Task for implementing chat bots, loading files, RAG and extracting form fields from natural language.

[dmitryrichard]

this is called metaprogramming brother

you are just doing it the harder way

you can do this easier by shot_firing

reflexing

or creating your own DSL

the workflow you imagine - looks alot like this - command sent

product aquired

( i edited my own code to display the reflex in action )

as a reflex the system detected a error ( the same one i introduced without me having to do anything)

in my case - CAPS is my own programming language - or rather a DSL - wherein i dont have to type commands to program anymore - i can type something as simple as

YOU directly mentined this

The output of any task can be accessed by subsequent tasks using {<task_name>} placeholders. A task can be python code instead of LLM instructions.

directly it looks like this

metaprogramming - bonus points if you have it immediatly implemented into the system ( which as you can see - even tho i didnt correct the error i introduced the system self healed itself to do that )

hope this helps

you also mentioned a python class library

connected to codes

alot of people on the forums do this - its just metaprogramming glad you found a way to introduce it brother !