document RASA_DUCKLING_HTTP_URL env variable usage

Rasa version: 1.10.7

Python version: 3.7

Operating system (windows, osx, …): Windows 10 1909

Issue: It seems like the Duckling-URL becomes a hard-coded part of the NLU-Model. Changing the Duckling-URL after training seems to have no effect.

Repro:

1. Train a model using f. e. http://localhost:8000 as Duckling-URL
2. Change the URL to f. e. http://localhost:55555 (but DONT retrain the model)
3. Start the Duckling-server on http://localhost:55555
4. Run rasa run
5. See failure after parsing any message

See also: https://forum.rasa.com/t/does-the-duckling-url-become-a-hard-coded-part-of-the-trained-model/31001

Error (including full traceback):

Failed to connect to duckling http server. Make sure the duckling server is running/healthy/not stale and the proper host and port are set in the configuration. More information on how to run the server can be found on github: https://github.com/facebook/duckling#quickstart Error: HTTPConnectionPool(host=‘localhost’, port=8000): Max retries exceeded with url: /parse (Caused by NewConnectionError(‘<urllib3.connection.HTTPConnection object at 0x000002BE546C5D48>: Failed to establish a new connection: <span class="error">[WinError 10061]</span> Es konnte keine Verbindung hergestellt werden, da der Zielcomputer die Verbindung verweigerte’))

Note: Port 8000 is still used.

Content of configuration file (config.yml) (if relevant): Before

language: de
 pipeline:

*   name: WhitespaceTokenizer
*   name: RegexFeaturizer
*   name: LexicalSyntacticFeaturizer
*   name: CountVectorsFeaturizer
*   name: CountVectorsFeaturizer
     analyzer: "char_wb"
     min_ngram: 1
     max_ngram: 4
*   name: DIETClassifier
     epochs: 100
*   name: EntitySynonymMapper
*   name: ResponseSelector
     epochs: 100
*   name: "DucklingHTTPExtractor"
     url: "http://localhost:8000"
     dimensions: <span class="error">["time"]</span>
     locale: "de_DE"
     timezone: "UTC"
     timeout : 3

After

language: de
 pipeline:

*   name: WhitespaceTokenizer
*   name: RegexFeaturizer
*   name: LexicalSyntacticFeaturizer
*   name: CountVectorsFeaturizer
*   name: CountVectorsFeaturizer
     analyzer: "char_wb"
     min_ngram: 1
     max_ngram: 4
*   name: DIETClassifier
     epochs: 100
*   name: EntitySynonymMapper
*   name: ResponseSelector
     epochs: 100
*   name: "DucklingHTTPExtractor"
     url: "http://localhost:55555" #This is the only change
     dimensions: <span class="error">["time"]</span>
     locale: "de_DE"
     timezone: "UTC"
     timeout : 3

[wochinge]

Exalate commented:

wochinge commented:

You could use an environment variable instead of a hardcoded URL in your configuration, e.g.

language: de
 pipeline:

*   name: WhitespaceTokenizer
*   name: RegexFeaturizer
*   name: LexicalSyntacticFeaturizer
*   name: CountVectorsFeaturizer
*   name: CountVectorsFeaturizer
     analyzer: "char_wb"
     min_ngram: 1
     max_ngram: 4
*   name: DIETClassifier
     epochs: 100
*   name: EntitySynonymMapper
*   name: ResponseSelector
     epochs: 100
*   name: "DucklingHTTPExtractor"
     url: $ {DUCKLING_URL} 

     #This is the only change
     dimensions: <span class="error">["time"]</span>
     locale: "de_DE"
     timezone: "UTC"
     timeout : 3

I see that this isn’t ideal in terms of usability, but otherwise we’d need some way to pass in the duckling url during loading the model. I think using an env variable still seems the best way to do this. Wdyt?

[N-Olbert]

Exalate commented:

N-Olbert commented:

According to the forum, this is defined functionality (see https://forum.rasa.com/t/does-the-duckling-url-become-a-hard-coded-part-of-the-trained-model/31001), so please consider this as an enhancement request rather than a bug.

Not being able to change the port/uri after training is bad because if a customer has already used the Duckling-port which was provided during training there is no way to change this port except to train a new model (which takes some time).