Snips NLU Server

This server allows you to run Snips NLU, an open source natural language understanding (NLU) engine, on your own premises.

Using this server is recommended if you want to build Jovo apps using the Snips NLU integration. It even features the ability to dynamically train the language model called dynamic entities.

Setup

To be able to interact with the Snips NLU engine, the server is implemented in Python. If you haven't already installed Python on your system, you can follow this guide. There have been some issues with using Python > 3.9, so we recommend sticking to either Python 2.7 or Python >= 3.5 <= 3.9.

After you've successfully configured your local environment, you can go ahead and download this repository:

# HTTPS
$ git clone https://github.com/jovotech/snips-nlu-server.git

# Alternative: SSH
$ git clone [email protected]:jovotech/snips-nlu-server.git

# Alternative: GitHub CLI
$ gh repo clone jovotech/snips-nlu-server

To manage dependencies on a project-scoped level, you need to create a virtual environment and activate it:

$ cd snips-nlu-server

# Create a virtual environment
$ python -m venv venv/

# Activate the environment
$ source venv/bin/activate

Depending on your shell you might need to run another script, read more here

To deactivate/exit the virtual environment, run:

$ deactivate

Install Dependencies

To be able to use such libraries as snips-nlu and flask, you'll need to install all required dependencies specified in requirements.txt:

$ pip install -r requirements.txt

If you get an error stating that the Rust compiler can't be found, try installing it using this guide.

Since the server uses @jovotech/model-snips for conversion, you need to install Node dependencies as well:

$ npm install

Additionally, Snips NLU requires you to download your language resources manually:

$ python -m snips-nlu download <language>

# or natively
$ snips-nlu download <language>

Run your Server

Set environment variables so Flask can find your server file:

# Linux/MacOS
$ export FLASK_APP=server/__init__.py

# Windows
$ set FLASK_APP=server/__init__.py

If you want to restart the server on changes automatically, add this to your environment:

# Linux/MacOS
$ export FLASK_ENV=development

# Windows
$ set FLASK_ENV=development

Start the server using the following command:

$ python -m flask run

# or natively
$ flask run

Training an Engine

The Snips NLU server provides an endpoint /engine/train that lets you train and persist a Snips NLU engine from a Jovo Language Model. It accepts the following query parameters:

• locale: The locale of the language model
• engine_id: The ID which is used to persist the NLU engine

You can choose between providing the model directly in the request body as part of the POST request, or set an environment variable MODEL_LOCATION, which holds the endpoint to the containing directory of your model (e.g. S3). Note that this endpoint is joined with the locale provided in the request query, so the final model endpoint will look something like this: ${MODEL_LOCATION}/${locale}.json.

The language model is then converted to a native Snips NLU dataset format, which can be further used to train the engine you'll later use to parse messages through. Using the provided query parameters engine_id and locale, this engine is then persisted locally to be reloaded when needed. This approach allows you to load your engine on demand, rather than to keep it running in a dedicated server instance. It also enables you to make use of Dynamic Entities, which will be explained below.

Dynamic Entities

By persisting the trained engines, it's possible to train a dedicated engine for dynamic entities on the fly and parse a message through it, if fit, otherwise the default engine is used. Note that in contrast to training a regular engine, this endpoint is called automatically by the Snips NLU integration, if you set entities in your output object. Learn more here.

The POST endpoint /engine/train/dynamic-entities accepts the following query parameters:

• locale: The locale of the language model
• engine_id: The ID which is used to persist the NLU engine
• entity: The dynamic entity name
• session_id: The current session ID

Apart from engine_id, which will be taken from configuration, these parameters will be set automatically. You can also set whether you want to pass the model automatically by setting passModels to true in the app configuration or provide your own models as described above.

Note that if you provide your own model, it only needs to contain the dynamic entity, along with every updated intent containing the entity.

Parsing a Message

The main endpoint of the server is /engine/parse, which will be called by the Snips NLU integration to parse a message and return the resolved values. It also includes parsing messages with dynamic entities. It accepts the following query parameters:

• engine_id: The ID which is used to load the NLU engine
• session_id: The current session ID
• locale: The locale of the current Jovo request

Similarly to the other endpoints, those parameters are either taken from the app configuration or set automatically, depending on the incoming request. The Snips NLU server then loads the previously trained and persisted engine and parses the message provided in the POST request body through it. The server will either use an engine trained for dynamic entities, if present, or fall back to the default engine.