Warning: This document is for an old version of Rasa NLU. The latest version is 0.15.1.

Using Rasa NLU as a HTTP server


Before you can use the server, you should train a model! See Training a New Model for your Project

The HTTP api exists to make it easy for non-python projects to use Rasa NLU, and to make it trivial for projects currently using wit/LUIS/Dialogflow to try it out.

Running the server

You can run a simple http server that handles requests using your projects with :

$ python -m rasa_nlu.server --path projects

The server will look for existing projects under the folder defined by the path parameter. By default a project will load the latest trained model.


Rasa NLU can ‘emulate’ any of these three services by making the /parse endpoint compatible with your existing code. To activate this, either add 'emulate' : 'luis' to your config file or run the server with -e luis. For example, if you would normally send your text to be parsed to LUIS, you would make a GET request to


in luis emulation mode you can call Rasa by just sending this request to


any extra query params are ignored by rasa, so you can safely send them along.

To use the emulation, pass the emulation mode to the server script:

$ python -m rasa_nlu.server --path projects --emulate wit


POST /parse (no emulation)

You must POST data in this format '{"q":"<your text to parse>"}', you can do this with

$ curl -XPOST localhost:5000/parse -d '{"q":"hello there"}'

By default, when the project is not specified in the query, the "default" one will be used. You can (should) specify the project you want to use in your query :

$ curl -XPOST localhost:5000/parse -d '{"q":"hello there", "project": "my_restaurant_search_bot"}'

By default the latest trained model for the project will be loaded. You can also query against a specific model for a project :

$ curl -XPOST localhost:5000/parse -d '{"q":"hello there", "project": "my_restaurant_search_bot", "model": "<model_XXXXXX>"}'

POST /train

You can post your training data to this endpoint to train a new model for a project. This request will wait for the server answer: either the model was trained successfully or the training exited with an error. Using the HTTP server, you must specify the project you want to train a new model for to be able to use it during parse requests later on : /train?project=my_project. The configuration of the model should be posted as the content of the request:

Using training data in json format:

language: "en"

pipeline: "spacy_sklearn"

# data contains the same json, as described in the training data section
data: {
  "rasa_nlu_data": {
    "common_examples": [
        "text": "hey",
        "intent": "greet",
        "entities": []

Using training data in md format:

language: "en"

pipeline: "spacy_sklearn"

# data contains the same md, as described in the training data section
data: |
  ## intent:affirm
  - yes
  - yep

  ## intent:goodbye
  - bye
  - goodbye

Here is an example request showcasing how to send the config to the server to start the training:

$ curl -XPOST -H "Content-Type: application/x-yml" localhost:5000/train?project=my_project \
    -d @sample_configs/config_train_server_md.yml


You cannot send a training request for a project already training a new model (see below).


The server will automatically generate a name for the trained model. If you want to set the name yourself, call the endpoint using localhost:5000/train?project=my_project&model=my_model_name

POST /evaluate

You can use this endpoint to evaluate data on a model. The query string takes the project (required) and a model (optional). You must specify the project in which the model is located. N.b. if you don’t specify a model, the latest one will be selected. This endpoint returns some common sklearn evaluation metrics (accuracy, f1 score, precision, as well as a summary report).

$ curl -XPOST localhost:5000/evaluate?project=my_project&model=model_XXXXXX -d @data/examples/rasa/demo-rasa.json | python -mjson.tool

    "accuracy": 0.19047619047619047,
    "f1_score": 0.06095238095238095,
    "precision": 0.036281179138321996,
    "predictions": [
            "intent": "greet",
            "predicted": "greet",
            "text": "hey",
            "confidence": 1.0
    "report": ...

GET /status

This returns all the currently available projects, their status (training or ready) and their models loaded in memory. also returns a list of available projects the server can use to fulfill /parse requests.

$ curl localhost:5000/status | python -mjson.tool

  "available_projects": {
    "my_restaurant_search_bot" : {
      "status" : "ready",
      "available_models" : [

GET /version

This will return the current version of the Rasa NLU instance.

$ curl localhost:5000/version | python -mjson.tool
  "version" : "0.8.2"

GET /config

This will return the default model configuration of the Rasa NLU instance.

$ curl localhost:5000/config | python -mjson.tool
    "config": "/app/rasa_shared/config_mitie.json",
    "data": "/app/rasa_nlu/data/examples/rasa/demo-rasa.json",
    "duckling_dimensions": null,
    "emulate": null,

DELETE /models

This will unload a model from the server memory

$ curl -X DELETE localhost:5000/models -d '{"project": "my_restaurant_search_bot", "model": <model_XXXXXX>}'


To protect your server, you can specify a token in your Rasa NLU configuration, e.g. by adding "token" : "12345" to your config file, or by setting the RASA_TOKEN environment variable. If set, this token must be passed as a query parameter in all requests, e.g. :

$ curl localhost:5000/status?token=12345

On default CORS (cross-origin resource sharing) calls are not allowed. If you want to call your Rasa NLU server from another domain (for example from a training web UI) then you can whitelist that domain by adding it to the config value cors_origin.

Serving Multiple Apps

Depending on your choice of backend, Rasa NLU can use quite a lot of memory. So if you are serving multiple models in production, you want to serve these from the same process & avoid duplicating the memory load.


Although this saves the backend from loading the same backend twice, it still needs to load one set of word vectors (which make up most of the memory consumption) per language and backend.

As stated previously, Rasa NLU naturally handles serving multiple apps : by default the server will load all projects found under the path directory defined in the configuration. The file structure under path directory is as follows :

- <path>
 - <project_A>
  - <model_XXXXXX>
  - <model_XXXXXX>
 - <project_B>
  - <model_XXXXXX>

So you can specify which one to use in your /parse requests:

$ curl 'localhost:5000/parse?q=hello&project=my_restaurant_search_bot'


$ curl -XPOST localhost:5000/parse -d '{"q":"I am looking for Chinese food", "project":"my_restaurant_search_bot"}'

You can also specify the model you want to use for a given project, the default used being the latest trained :

$ curl -XPOST localhost:5000/parse -d '{"q":"I am looking for Chinese food", "project":"my_restaurant_search_bot", "model":<model_XXXXXX>}'

If no project is to be found by the server under the path directory, a "default" one will be used, using a simple fallback model.

Server Parameters

There are a number of parameters you can pass when running the server.

$ python -m rasa_nlu.server

Here is a quick overview:

usage: server.py [-h] [-e {wit,luis,dialogflow}] [-P PORT]
                 [--pre_load PRE_LOAD [PRE_LOAD ...]] [-t TOKEN] [-w WRITE]
                 --path PATH [--cors [CORS [CORS ...]]]
                 [--max_training_processes MAX_TRAINING_PROCESSES]
                 [--num_threads NUM_THREADS] [--endpoints ENDPOINTS]
                 [--wait_time_between_pulls WAIT_TIME_BETWEEN_PULLS]
                 [--response_log RESPONSE_LOG] [--storage STORAGE] [-c CONFIG]
                 [--debug] [-v]

parse incoming text

optional arguments:
  -h, --help            show this help message and exit
  -e {wit,luis,dialogflow}, --emulate {wit,luis,dialogflow}
                        which service to emulate (default: None i.e. use
                        simple built in format)
  -P PORT, --port PORT  port on which to run server
  --pre_load PRE_LOAD [PRE_LOAD ...]
                        Preload models into memory before starting the server.
                        If given `all` as input all the models will be loaded.
                        Else you can specify a list of specific project names.
                        Eg: python -m rasa_nlu.server --pre_load project1
                        --path projects -c config.yaml
  -t TOKEN, --token TOKEN
                        auth token. If set, reject requests which don't
                        provide this token as a query parameter
  -w WRITE, --write WRITE
                        file where logs will be saved
  --path PATH           working directory of the server. Models areloaded from
                        this directory and trained models will be saved here.
  --cors [CORS [CORS ...]]
                        List of domain patterns from where CORS (cross-origin
                        resource sharing) calls are allowed. The default value
                        is `[]` which forbids all CORS requests.
  --max_training_processes MAX_TRAINING_PROCESSES
                        Number of processes used to handle training requests.
                        Increasing this value will have a great impact on
                        memory usage. It is recommended to keep the default
  --num_threads NUM_THREADS
                        Number of parallel threads to use for handling parse
  --endpoints ENDPOINTS
                        Configuration file for the model server as a yaml file
  --wait_time_between_pulls WAIT_TIME_BETWEEN_PULLS
                        Wait time in seconds between NLU model serverqueries.
  --response_log RESPONSE_LOG
                        Directory where logs will be saved (containing queries
                        and responses).If set to ``null`` logging will be
  --storage STORAGE     Set the remote location where models are stored. E.g.
                        on AWS. If nothing is configured, the server will only
                        serve the models that are on disk in the configured
  -c CONFIG, --config CONFIG
                        Default model configuration file used for training.
  --debug               Print lots of debugging statements. Sets logging level
                        to DEBUG
  -v, --verbose         Be verbose. Sets logging level to INFO