Warning: This document is for an old version of rasa NLU.

Migrating an existing app

rasa NLU is designed to make migrating from wit/LUIS/api.ai as simple as possible. The TLDR instructions for migrating are:

Banana Peels

Just some specific things to watch out for for each of the services you might want to migrate from


Wit used to handle intents natively. Now they are somewhat obfuscated. To create an intent in wit you have to create and entity which spans the entire text. The file you want from your download is called expressions.json


When you download your model, the entity locations are specified by the index of the tokens. This is pretty fragile because not every tokenizer will behave the same as LUIS’s, so your entities may be incorrectly labelled. Run your training once and you’ll get a copy of your training data in the model_XXXXX dir. Do any fixes required and use that to train. Use the visualizer (see Visualizing the Training Data) to spot mistakes easily.

To specify the tokenizer you are using you need to define the luis_data_tokenizer configuration variable.


api app exports generate multiple files rather than just one. Put them all in a directory (see data/examples/api in the repo) and pass that path to the trainer.


To make rasa NLU easy to try out with existing projects, the server can emulate wit, LUIS, or api.ai. In native mode, a request / response looks like this :

$ curl -XPOST localhost:5000/parse -d '{"q":"I am looking for Chinese food"}' | python -mjson.tool
  "text": "I am looking for Chinese food",
  "intent": "restaurant_search",
  "confidence": 0.4794813722432127,
  "entities": [
      "start": 17,
      "end": 24,
      "value": "chinese",
      "entity": "cuisine"

if we run in wit mode (e.g. python -m rasa_nlu.server -e wit)

then instead have to make a GET request

$ curl 'localhost:5000/parse?q=hello' | python -mjson.tool
        "_text": "hello",
        "confidence": 0.4794813722432127,
        "entities": {},
        "intent": "greet"

similarly for LUIS, but with a slightly different response format

$ curl 'localhost:5000/parse?q=hello' | python -mjson.tool
    "entities": [],
    "query": "hello",
    "topScoringIntent": {
        "intent": "inform",
        "score": 0.4794813722432127

and finally for api.ai

$ curl 'localhost:5000/parse?q=hello' | python -mjson.tool
    "id": "ffd7ede3-b62f-11e6-b292-98fe944ee8c2",
    "result": {
        "action": null,
        "actionIncomplete": null,
        "contexts": [],
        "fulfillment": {},
        "metadata": {
            "intentId": "ffdbd6f3-b62f-11e6-8504-98fe944ee8c2",
            "intentName": "greet",
            "webhookUsed": "false"
        "parameters": {},
        "resolvedQuery": "hello",
        "score": null,
        "source": "agent"
    "sessionId": "ffdbd814-b62f-11e6-93b2-98fe944ee8c2",
    "status": {
        "code": 200,
        "errorType": "success"
    "timestamp": "2016-11-29T12:33:15.369411"