Rasa API

[wochinge]

Suggested API commands:

a python API to use the Rasa tools within e.g. ipython notebooks

Goal

• don't expose internal stuff (agent, file loading etc)
• be consistent between core / nlu

Changes to Rasa Stack

import rasa

rasa.train(config=..., data=...) # train core + nlu
rasa.run(config=...) # run core + nlu
rasa.test(config=..., data=...) # test core + nlu

rasa.model.load(path=...) -> Model # load core + nlu model
rasa.model.persist(core=..., nlu=..., path=...) # persist core + nlu model

rasa.data.load(...) # load core + nlu data, return tuple
rasa.data.load_stories(...) # load core data
rasa.data.load_messages(...) # load nlu data

rasa.config.load(domain=..., config=..., credentials=..., endpoints=...)

rasa.jupyter.chat() # chat with the bot within in a jupyter notebook

Changes to Core and NLU

Modules

• rename rasa_nlu.run to rasa_nlu.shell
• rename rasa_nlu.server to rasa_nlu.run
• rename evaluate scripts to test ✅
• add a module rasa.remote which covers all the remote stuff
• add a module rasa.local which covers the interaction with the local server

Training

• rename rasa_nlu.train.do_train() to rasa_nlu.train.train() ✅
• pass filepath to rasa_nlu.train.do_train() instead of file ✅
• rename rasa_core.train.train_dialogue_model() to rasa_core.train.train() ✅

Testing

• create method rasa_nlu.test.cross_validate() which can be called with file paths etc and creates the agent inside ✅
• create method rasa_core.test.test() which can be called with file paths etc and creates the agent inside ✅
• create method rasa_core.test.compare() (compare multiple policies) ✅

Resulting Jupyter notebooks

Both should be possible: Demonstrating Rasa Core, NLU separately and showing them as one

import rasa_nlu

nlu_data = "..."
nlu_config = "..."

rasa_nlu.train(...)
rasa_nlu.test(...)

import rasa_core

stories = "..."
domain = "..."
policies = "..."
rasa_core.train(...)
rasa_core.test()

import rasa
rasa.jupyter.chat()

import rasa

nlu_data = "..."
stories = "..."
domain = "..."
config = "..."

rasa.train(...)
rasa.test(...)
rasa.jupyter.chat(...)

[MetcalfeTom]

Since you want to implement rasa.train() and rasa.test(), is it also possible to rename these other functions? rasa_nlu.train() would be a lot better than rasa_nlu.train.train() and same with test()

[tmbo]

Considering the scripts in NLU:

• the current "run" (which starts a shell to try out nlu messages) script should be comparable to the core cmdline mode
• server.py should just contain the server

There are a couple more concepts that we should support:

• loading data
• persisting / loading a model

[wochinge]

• the current "run" (which starts a shell to try out nlu messages) script should be comparable to the core cmdline mode

So you suggest to leave the module names as they are?

[tmbo]

well, not sure whatever you think makes it most consistent

[wochinge]

Updated the issue according to Tom's suggestion:

• added rasa.data module
• added rasa.local module
• added rasa.model module
• decided not to rename the nlu modules run and server since this would lead to a lot of breaking things

[howl-anderson]

Nice Job! Few comments about it:

Maybe we can add features to Rasa Python API to make it more easily integrated with (independent from) service container (HTTP server, RPC (grpc / thrift / protobf) server) or directly used in python interpreter (such as python interaction session and jupyter notebook). ideally, it look like this:

import rasa

# nlu_model can be local data object, agent to remote service or dummy object for debug
nlu_model = rasa.model.nlu.load(path='...')  # load NLU model from local

core_model = rasa.model.core.load(path='...', nlu_entry=nlu_model)

# in python, user can use rasa by
result = core_model.eval('I want book a hotel in LA tomorrow')
print(result)  # output: OK. I booked xxx hotel for you
print(core_model.current_states)  # view current internal states, very useful for real time debug
print(nlu_model.current_states)  # view current internal states, very useful for real time debug

# in jupyter notebook, user can use rasa by
jupyter_wrapper = rasa.jupyter.wrap_core(core_entry=core_model)
result = jupyter_wrapper.eval('I want book a hotel in LA tomorrow')
print(result)  # output jupyter friendly message: OK. I booked xxx hotel for you
print(jupyter_wrapper.current_states)  # view current internal states, display friendly for jupyter

# wrap it for start a http service
http_server = rasa.service.http(core_entry=core_model)
http_server.serve_forever(port='80', host='0.0.0.0')

# wrap it for start a grpc service
grpc_server = rasa.service.grpc(core_entry=core_model)
grpc_server.serve_forever(port='9090', host='0.0.0.0')

[suenpun]

great! I am using gRPC now , realy need this RPC features! Looking forward to it!@howl-anderson @tmbo