Skip to main content

Python bindings for Snips Hermes Protocol

Project description

Hermes Python


The hermes-python library provides python bindings for the Hermes protocol that snips components use to communicate together over MQTT. hermes-python allows you to interface seamlessly with the Snips platform and kickstart development of Voice applications.

hermes-python abstracts away the connection to the MQTT bus and the parsing of incoming and outcoming messages from and to the components of the snips platform.


Pre-compiled wheels are available for Python 2.7+ and Python 3.5

The pre-compiled wheels supports the following platform tags :

  • manylinux1_x86_64

  • armv7l, armv6

  • macos

If you want to install hermes-python on another platform, you have to build it from source.


The library is packaged as a pre-compiled platform wheel, available on PyPi.

It can be installed with : pip install hermes-python.

Or you can add it to your requirements.txt file.

Building from source

If you want to use hermes-python on platforms that are not supported, you have to manually compile the wheel.

You need to have rust and cargo installed :

curl -sSf

Clone, the hermes-protocol repository :

git clone
cd hermes-protocol/platforms/hermes-python

You can then build the wheel :

virtualenv env
source env/bin/activate
python bdist_wheel

The built wheels should be in platforms/hermes-python/dist

You can install those with pip : pip install platforms/hermes-python/dist/<your_wheel>.whl

Advanced wheel building

We define a new API for including pre-compiled shared objects when building a platform wheel.

python bdist_wheel

This command will compile the hermes-mqtt-ffi Rust extension, copy them to an appropriate location, and include them in the wheel.

We introduce a new command-line argument : include-extension which is a way to include an already compiled (in previous steps) hermes-mqtt-ffi extension in the wheel.

Its usage is the following : include-extension=<default | the/path/to/your/extension.[so|dylib]>

For instance :

python bdist_wheel --include-extension=default

The default value for include-extension will look up for pre-compiled extension in the default paths (in hermes-protocol/target/release/libhermes_mqtt_ffi.[dylib|so] and hermes-protocol/platforms/hermes-python/hermes_python/dylib).

python bdist_wheel --include-extension=<the/path/to/your/extension.[so|dylib]>

When doing x-compilation, you can also specify the target platform :

python bdist_wheel --include-extension=<the/path/to/your/extension.[so|dylib]> --plat-name=<the_platform_tag>


The lifecycle of a script using hermes-python has the following steps :

  • Initiating a connection to the MQTT broker

  • Registering callback functions to handle incoming intent parsed by

    the snips platform

  • Listening to incoming intents

  • Closing the connection

Let’s quickly dive into an example :

Let’s write an app for a Weather Assistant ! This code implies that you created a weather assistant using the Snips Console, and that it has a searchWeatherForecast intent. Or you could download this weather Assistant .

from import Hermes

MQTT_ADDR = "localhost:1883"        # Specify host and port for the MQTT broker

def subscribe_weather_forecast_callback(hermes, intent_message):    # Defining callback functions to handle an intent that asks for the weather.
    print("Parsed intent : {}".format(intent_message.intent.intent_name))

with Hermes(MQTT_ADDR) as h: # Initialization of a connection to the MQTT broker
    h.subscribe_intent("searchWeatherForecast", subscribe_weather_forecast_callback) \  # Registering callback functions to handle the searchWeatherForecast intent
    # We get out of the with block, which closes and releases the connection.

This app is a bit limited as it only prints out which intent was detected by our assistant. Let’s add more features.

Handling the IntentMessage object

In the previous example, we registered a callback that had this signature.

subscribe_intent_callback(hermes, intent_message)

The intent_message object contains information that was extracted from the spoken sentence.

For instance, in the previous code snippet, we extracted the name of the recognized intent with


We could also retrieve the associated confidence score the NLU engine had when classifying this intent with


Extracting slots

Here are some best practices when dealing with slots. The IntentMessage object has a slots attribute.

This slots attributes is a container that is empty when the intent message doesn’t have slots :

assert len(intent_message.slots) == 0

This container is a dictionary where the key is the name of the slot, and the value is a list of all the slot values for this slot name.

You can access these values in two ways :

assert len(intent_message.slots.slot1) == 0
assert len(intent_message.slots["slot1"]) == 0

The slot values are of type NluSlot which is a deeply nested object, we offer convenience methods to rapidly access the slot_value attribute of the NluSlot.

To access the first slot_value of a slot called myslot, you can use :


You can also access all the slot_value of a slot called myslot :


Let’s add to our Weather assistant example.

We assume that the searchWeatherForecast has one slot called forecast_location, that indicates which location the user would like to know the weather at.

Let’s print all the forecast_location slots :

for slot in intent_message.slots.forecast_location:
    name = slot.slot_name
    confidence = slot.confidence_score
    print("For slot : {}, the confidence is : {}".format(name, confidence))

The dot notation was used, but we can also use the dictionary notation :

for slot in intent_message.slots.forecast_location:
    name = slot["slot_name"]

Some convenience methods are available to easily retrieve slot values :

Retrieving the first slot value for a given slot name

slot_value = intent_message.slots.forecast_location.first()

Retrieving all slot values for a given slot name

slot_values = intent_message.slots.forecast_location.all()

Coming back to our example, we can now have the app print the forecast_location slot value back to the user :

def subscribe_weather_forecast_callback(hermes, intent_message):
    slot_value = intent_message.slots.forecast_location.first().value
    print("The slot was : {}".format(slot_value)

Managing sessions

The Snips platform includes support for conversations with back and forth communication between the Dialogue Manager and the client code. Within the Snips platform, a conversation happening between a user and her assistant is called a session.

In this document, we will go through the details of how to start, continue and end a session.

In its default setup, you initiate a conversation with your assistant by pronouncing the defined wake-word. You say your request out-loud, an intent is extracted from your request, and triggers the portion of the action code you registered to react to this intent. Under the hood, the Dialogue Manager starts a new session when the wake-word is detected. The session is then ended by the action code.

Starting a session

A session can be also be started programmatically. When you initiate a new session, the Dialogue Manager will start the session by asking the TTS to say the text (if any) and wait for the answer of the end user.

You can start a session in two manners :

  • with an action

  • with a notification

When initiating a new session with an action, it means the action code will expect a response from the end user.

For instance: You could have an assistant that books concerts tickets for you. The action code would start a session with an action, having the assistant asking for what band you would like to see live.

When initiating a new session with a notification, it means the action code only inform the user of something without expecting a response.

For instance: Instead of pronouncing your defined wake-word, you could program a button to initiate a new session.

Let’s build up on our previous example of an assistant that book concerts tickets for you. Here, we are going to initiate a new session with an action, filtering on the intent the end-user can respond with.

from import Hermes, MqttOptions

with Hermes(mqtt_options=MqttOptions()) as h:
        "What band would you like to see live ?",
        True, False, None)

Let’s say that we added a physical button to initiate a conversation with our concert tickets booking assistant. We could use this button to initiate a new session and start talking immediately after pressing the button instead of relying on triggering a wake-word.

When the button is pressed, the following code could be ran :

hermes.publish_start_session_notification("office", None, None)

This would initiate a new session on the office site id.

Ending a session

To put an end to the current interaction the action code can terminate a started session. You can optionally terminate a session with a session with a message that should be said out loud by the TTS.

Let’s get back to our concert tickets booking assistant, we would end a session like this :

from import Hermes, MqttOptions

def find_shows(band):

def findLiveBandHandler(hermes, intent_message):
    band =
    shows = find_shows(band)
    hermes.publish_end_session(intent_message.session_id, "I found {} shows for this band !".format(len(shows)))

with Hermes(mqtt_options=MqttOptions()) as h:
        .subscribe_intent("findLiveBand", findLiveBandHandler)\

Continuing a session

You can programmatically extend the lifespan of a dialogue session, expecting interactions from the end users. The typical use of continuing a session is for your assistant to ask additional information to the end user.

Let’s continue with our concert tickets booking assistant, after starting a session, we will continue a session, expecting the user to tell us how many tickets the assistant should buy.

import json
from import Hermes, MqttOptions

required_slots = {  # We are expecting these slots.
    "band": None,
    "number_of_tickets": None

def ticketShoppingHandler(hermes, intent_message):
    available_slots = json.loads(intent_message.custom_data)

    band_slot = or available_slots["band"]
    number_of_tickets = intent_message.slots.number_of_tickets.first().value or available_slots["number_of_tickets"]

    available_slots["band"] = band_slot
    available_slots["number_of_tickets"] = number_of_tickets

    if not band_slot:
        return hermes.publish_continue_session(intent_message.session_id,
                                               "What band would you like to see live ?",

    if not number_of_tickets:
        return hermes.publish_continue_session(intent_message.session_id,
                                               "How many tickets should I buy ?",

    return hermes.publish_end_session(intent_message.session_id, "Ok ! Consider it booked !")

with Hermes(mqtt_options=MqttOptions("raspi-anthal-support.local")) as h:
        .subscribe_intent("ticketShopping", ticketShoppingHandler)\

Slot filling

You can programmatically continue a session, and asking for a specific slot. If we build on our previous example, we could continue a dialog session by specifying which slot the assistant expects from the end-user.

import json
from import Hermes, MqttOptions

required_slots_questions = {
    "band": "What band would you like to see live ?",
    "number_of_tickets": "How many tickets should I buy ?"

def ticketShoppingHandler(hermes, intent_message):
    available_slots = json.loads(intent_message.custom_data)

    band_slot = or available_slots["band"]
    number_of_tickets = intent_message.slots.number_of_tickets.first().value or available_slots["number_of_tickets"]

    available_slots["band"] = band_slot
    available_slots["number_of_tickets"] = number_of_tickets

    missing_slots = filter(lambda slot: slot is None, [band_slot, number_of_tickets])

    if len(missing_slots):
        missing_slot = missing_slots.pop()
        return hermes.publish_continue_session(intent_message.session_id,
        return hermes.publish_end_session(intent_message.session_id, "Ok ! Consider it booked !")

with Hermes(mqtt_options=MqttOptions("raspi-anthal-support.local")) as h:
        .subscribe_intent("ticketShopping", ticketShoppingHandler)\

Dynamic Vocabulary using Entities Injection

Please refer to the official documentation for further information.

Sometimes, you want to extend your voice assistant with new vocabulary it hasn’t seen when it was trained. For instance, let’s say that you have a bookstore voice assistant, that you update every week with new book titles that came out.

The snips platform comes with the Entities Injection feature, which allows you to update both the ASR and the NLU models directly on the device to understand new vocabulary.

Each intent within an assistant may contain some slots, and each slot has a specific type that we call an entity. If you have a book_title entity that contains a list of book titles in the inventory of your book store, Entities Injection lets you add new titles to this list.

To inject new entity values, you have multiple operations at your disposal :

  • add adds the list of values that you provide to the existing

    entity values.

  • addFromVanilla removes all the previously injected values to

    the entity, and then, adds the list of values provided. Note that the entity values coming from the console will be kept.

Let’s see how an injection would be performed by the action code :

from import Hermes
from hermes_python.ontology.injection import InjectionRequestMessage, AddInjectionRequest, AddFromVanillaInjectionRequest

def retrieve_new_book_releases():
    return ["The Half-Blood Prince", "The Deathly Hallows"]

def retrieve_book_inventory():
    return ["The Philosopher's Stone", "The Chamber of Secrets", "The Prisoner of Azkaban", "The Goblet of Fire",
            "The Order of the Phoenix", "The Half-Blood Prince", "The Deathly Hallows"]

# First example : We just add weekly releases

operations =  [
    AddInjectionRequest({"book_titles" : retrieve_new_book_releases() }),

request1 = InjectionRequestMessage(operations)

with Hermes("localhost:1883") as h:

# Second example : We reset all the previously injected values of the book_title entity, and then, adds the list of values provided

operations =  [
    AddInjectionRequest({"book_titles" : retrieve_book_inventory() }),

request2 = InjectionRequestMessage(operations)

with Hermes("localhost:1883") as h:

Careful, performing an entity injection is a CPU and memory intensive task. You should not trigger multiple injection tasks at the same time on devices with limited computing power.

You can register a callback so that your code knows when an injection process is completed :

def injection_completed(hermes, injection_complete_message):
    print("The injection operation with id {} completed !".format(injection_complete_message.request_id))

with Hermes("localhost:1883") as h:

You can monitor the progress of your injection request with snips-watch -vvv.

You can also reset the injected vocabulary of your assistant to its factory settings using the request_injection_reset` method of ``hermes. Since the operation of resetting the injection is asynchronous, you can register a callback to know when the injection reset process is completed :

def injection_reset_completed(hermes, injection_reset_complete_message):
    print("The injection reset operation with id {} completed !".format(injection_reset_complete_message.request_id))

with Hermes("localhost:1883") as h:

Configuring MQTT options

The connection to your MQTT broker can be configured with the hermes_python.ffi.utils.MqttOptions class.

The Hermes client uses the options specified in the MqttOptions class when establishing the connection to the MQTT broker.

Here is a code example :

from import Hermes
from hermes_python.ffi.utils import MqttOptions

mqtt_opts = MqttOptions()

def simple_intent_callback(hermes, intent_message):
    print("I received an intent !")

with Hermes(mqtt_options=mqtt_opts) as h:

Here are the options you can specify in the MqttOptions class :

  • broker_address: The address of the MQTT broker. It should be

    formatted as ip:port.

  • username: Username to use on the broker. Nullable

  • password: Password to use on the broker. Nullable

  • tls_hostname: Hostname to use for the TLS configuration.

    Nullable, setting a value enables TLS

  • tls_ca_file: CA files to use if TLS is enabled. Nullable

  • tls_ca_path: CA path to use if TLS is enabled. Nullable

  • tls_client_key: Client key to use if TLS is enabled. Nullable

  • tls_client_cert: Client cert to use if TLS is enabled. Nullable

  • tls_disable_root_store: Boolean indicating if the root store

    should be disabled if TLS is enabled.

Let’s connect to an external MQTT broker that requires a username and a password :

from import Hermes
from hermes_python.ffi.utils import MqttOptions

mqtt_opts = MqttOptions(username="user1", password="password", broker_address="")

def simple_intent_callback(hermes, intent_message):
    print("I received an intent !")

with Hermes(mqtt_options=mqtt_opts) as h:

Configuring Dialogue

hermes-python offers the possibility to configure different aspects of the Dialogue system.

Enabling and disabling intents on the fly

It is possible to enable and disable intents of your assistant on the fly. Once an intent is disabled, it will not be recognized by the NLU.

Note that intents in the intent filters of started or continued session will take precedence over intents that are enabled/disabled in the configuration of the Dialogue.

You can disable/enable intents with the following methods :

from hermes_python.ontology.dialogue import DialogueConfiguration

dialogue_conf = DialogueConfiguration()                          \
                        .disable_intent("intent1")               \
                        .enable_intent("intent2")                \
                        .enable_intents(["intent1", "intent2"])  \
                        .disable_intents(["intent2", "intent1"])


Configuring Sound Feedback

Enabling and disabling sound feedback

By default, the snips platform notify the user of different events of its lifecycle with sound. It emits a sound when the wakeword is detected, or when the NLU engine (natural understanding engine) has successfuly extracted an intent from a spoken sentence.

hermes-python allows to disable this sound feedback programmatically, by sending a message to the snips platform, specifying the siteId where the sound feedback should be disabled.

from import Hermes
from import SiteMessage

with Hermes("localhost:1883") as h:

Making the TTS play custom sounds

The snips-platform allows you to register custom sounds which can be played later by the TTS engine.

hermes-python allows you to register sounds on the fly, by specifying a string identifier for the sound, and providing a wav file.

For instance, let’s say that your assistant tells a bad joke and that you want to play a ba dum tss sound at the end of the punchline.

from builtins import bytearray
from import Hermes
from hermes_python.ontology.tts import RegisterSoundMessage

# Step 1 : We read a wav file
def read_wav_data():
    with open('ba_dum_tss.wav', 'rb') as f:
        read_data =
    return bytearray(read_data)

# Step 2 : We register a sound that will be named "bad_joke"
sound = RegisterSoundMessage("bad_joke", read_wav_data())

def callback(hermes, intent_message):
    hermes.publish_end_session(intent_message.session_id, "A very bad joke ... [[sound:bad_joke]]")  # Step 4 : You play your registered sound

with Hermes("localhost:1883") as h:
        .register_sound(sound)\    # Step 3 : You register your custom sound

In the TTS string, when you specify the sound you want to play, you need to follow the syntax : [[sound:<your_sound_id>]]

Enabling Debugging

You can debug hermes-python if you encounter an issue and get a better stacktrace that you can send us.

To do so, you have to set the rust_logs_enabled flag to True when you create an instance of the Hermes class :

from import Hermes

def callback(hermes, intent_message):

with Hermes("localhost:1883", rust_logs_enabled=True) as h:
    h.subscribe_intent("...", callback)

You should then execute your script with the RUST_LOG environment variable : RUST_LOG=TRACE python

Release Checklist

Everytime you need to perform a release, do the following steps :

  • [ ] Commit all changes to the project for said release

  • [ ] Write all the changes introduced in the Changelog

    (source/HISTORY.rst file) and commit it

  • [ ] Run tests

  • [ ] Build the documentation and commit the README.rst

  • [ ] Bump the version and commit it

  • [ ] Upload to PyPI

Build details

Creating macOS wheels

The build script : build_scripts/ uses pyenv to generate hermes-python wheels for different versions of python.

To be able to run it, you need to :

  • install pyenvbrew install pyenv. Then follow the additional

    steps detailled

  • you then have to install python at different versions:

pyenv install --list to list the available version to install * Before installing and building the different python version from sources, install the required dependencies : Link here

That’s it ! History ========== 0.8.1 (2019-10-03) —————— * Hotfix : adding back DialogueConfiguration in the main module + Conversion function for SessionTermination object

0.8.0 (2019-09-10)

  • Adds subscription to injection lifecycle events : subscribe_injection_complete, subscribe_injection_reset_complete

  • Adds a component field to the SessionTerminationType class

  • Introduces alternatives intent resolutions

  • Fixes folder creation issue when building the wheel from sources

0.7.0 (2019-05-14)

  • Introduces Entities Injection API.

0.6.1 (2019-05-10)

  • Introduces register_sound API

0.5.2 (2019-05-07)

  • Fixes nullable fields in Dialogue ontology and brings more type annotations

0.5.1 (2019-05-06)

  • introduces new (cli) API to build python wheel that include pre-compiled hermes-mqtt-ffi extension.

0.5.0 (2019-04-19)

  • Adds APIs to enable and disable sound feedback.

0.4.1 (2019-03-29)

  • Re-enables debugging of hermes-python with the rust_logs_enabled flag

  • AmountOfMoneyValue, InstantTimeValue and DurationValue slot values now use Precision and Grain enumerations

0.4.0 (2019-03-20)

  • Adds support to configure the Dialogue Mananger : enabling and disabling intents on the fly.

  • Adds slot filling API : You can ask for a specific slot when continuing a session

  • adding support for OrdinalSlot

0.3.3 (2019-03-06)

  • Fixes a bug with publish_start_session_notification that didn’t take the text parameter into account.

0.3.2 (2019-02-25)

  • Fixes an important bug that gave the argument hermes the wrong type for every registered callback.

  • Fixes an important bug that caused the program to crash when parsing intentMessage that had no slots.

0.3.1 (2019-02-25)

  • Fixes import bug with templates, the hermes_python.ffi.utils module now re-exports MqttOptions

0.3.0 (2019-02-25)

  • IntentClassifierResult’s probability field has been renamed to confidence_score.

  • Introduces support for snips-platform 1.1.0 - 0.61.1.

0.2.0 (2019-02-04)

  • Introduces options to connect to the MQTT broker (auth + TLS are now supported).

0.1.29 (2019-01-29)

  • Fixes bug when deserializing TimeIntervalValue that used wrong encode method instead of decode.

0.1.28 (2019-01-14)

  • Fixes bug when the __exit__ method was called twice on the Hermes class.

  • Introduces two methods to the public api : connect and disconnect that should bring more flexibility

0.1.27 (2019-01-07)

  • Fixed broken API introduced in 0.1.26 with the publish_continue_session method of the Hermes class.

  • Cast any string that goes in the mqtt_server_adress parameter in the constructor of the Hermes class to be a 8-bit string.

0.1.26 (2019-01-02)

  • LICENSING : This wheel now has the same licenses as the parent project : APACHE-MIT.

  • Subscription to not recognized intent messages is added to the API. You can now write your own callbacks to handle unrecognized intents.

  • Adds send_intent_not_recognized flag to continue session : indicate whether the dialogue manager should handle non recognized intents by itself or sent them as an IntentNotRecognizedMessage for the client to handle.

0.1.25 (2018-12-13)

  • Better error handling : Errors from wrapped C library throw a LibException with detailled errors.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Built Distributions

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page