Is there a developer control, either in api.ai, or at Actions on Google side, to disable the <"I couldn't hear what you just said" + Action repeat> behavior that occurs whenever user does not respond quickly enough after an utterance?
It can bodge some Actions; either those with carefully controlled language interactions, or those that perform "once only" functions as they speak.
You need to implement support for no-input handling. If you don't explicitly handle it, then either the API.AI agent will use its defaults or the assistant platform will provide defaults.
For API.AI, follow these instructions: https://developers.google.com/actions/apiai/fulfillment#no-input_reprompting
You can take a look at the Number Genie sample for prompts that will prompt the user to respond: https://github.com/actions-on-google/apiai-number-genie-nodejs
For clarity, I'm summarizing Leon's advice and my own background research:
We cannot disable reprompting. That is, we cannot design interactions in which an Action utterance is not followed immediately by a user utterance (i.e., where user silence is part of the conversation dynamic).
Through Fulfillment, we can provide alternate prompt form(s).
But we cannot disable or alter any of these system mechanisms: prompt after a few seconds of silence, if no reply repeat, after three tries quit the Action.
Related
There have been triggering problems with my action since a few days ago. The queries that should have been handled by my action were routed to Google Assistant main flow. This happens on both Android Phone, and Google Home.
Steps to repro:
Speak: OK Google, talk to Tinker Doodle.
Assistant: Welcome to Tinker Doodle, what can I do for you?
Speak: Available commands.
Assistant: (Abruptly end Tinker Doodle conversation, list general commands on Assistant.)
I'd expect Assistant to stay in Tinker Doodle conversation, and feed the input to my action.
This makes Tinker Doodle almost unusable. Can you help with this?
I configured the NO_MATCH system intent to call my webhook, since I use my own NLP.
This worked well on Android Phone and Google Home, until a few days ago. There is no problem running in simulator on Action Builder.
Here are the screenshots of the main scene and NO_MATCH intent from Action Builder.
It isn't clear, but this sounds like it may be related to recent announcements that, in some cases, phrases that don't match a specific Intent may cause your Action to close so the Assistant can handle the phrase instead.
Even besides this, handling things with NO_MATCH is generally undesirable, since that will only happen three times in a row before the Action is forcibly closed.
Instead, you should create an Intent that can handle "any" input and route that input to your handler using this method. That involves:
Creating a new Type (I usually call it "Any") that accepts Free form text
Creating an Intent (which I have named "matchAny") that accepts values of this type through its training phrases (or even just one phrase that accepts a value of this type)
In your Scene, add this as an Intent that can be matched, and then set the handler for your webhook when it does.
Rather than using no_match, you can employ the design that the custom-nlu sample uses:
Have a 'Main' scene that tries to match on a user_utterance intent:
Then the user_utterance matches on everything using the any data type:
When you go to the Simulator, any query should match your intent explicitly and then, as part of the sample, it will echo your response:
In my conversation dialogflow, I would like to add some progress messages like hang in with me, I'm looking up for that data or similar in the conversation. Is there any guidance or best practice to do this?
Unfortunately, there is no good way to do this at this time. If your webhook takes longer than about 5 seconds, Dialogflow will return one of the default responses it is set with. If you're not using Dialogflow, the Action SDK will say your webhook isn't responding and will close the conversation.
There is currently no way to send a reply, and then send another reply without the user saying something first.
One workaround might be to have the default response be something like "I'm looking that information up. Ask me again in a few seconds." When your lookup finally completes, cache the information so when/if the user asks the question again, you can return it more quickly.
Depending how long it takes, you may also wish to register a dynamic reprompt. This will send an event to your webhook if the user doesn't say anything. In a situation like this, they may say nothing for a few seconds, but that may be long enough for you to have computed the reply. So after a few seconds of silence you can suddenly announce "I've figured it out, the answer you were looking for is..." or something similar. This has some limitations - you can only reprompt twice like this before Google sends you a final reprompt and closes the conversation.
Although the platform does support notifications, these are still in developer preview and don't work with all devices. They also don't quite continue the conversation (it doesn't just start talking) - they just send a notification to a phone that there is a message and that they can restart the conversation. Depending on your use case, this may be useful combined with the above.
Update
The Media Response includes a feature that we can take advantage of to handle this. Similar to the dynamic reprompt method above, you'll get a call automatically when the media you're playing ends. So you can play a short "hold music" and your webhook will be called when it is finished. You can then either give the result or say you're still working on it and play more hold music.
My program needs to react to the user not taking any action on a GNotification.
More specificially, a piece of data is written to the database only if the user does not press the "undo" button on the notification sent after the data's creation. My target deployment scenario does have notifications enabled and a real timeout value.
To be precise: Moving the notification "away" / deleting it should also count as such a timeout.
1) Is there a built-in way to 'listen' to notification timeouts?
2) If not, how could I still implement similar behavior?
I would use the D-Bus org.freedesktop.Notifications interface. Although it is still a draft specification, it does appear stable. My experience accessing the D-Bus interface using Vala has been that it is easier to use and gives the full feature set of the specification. GNotification doesn't seem to be as feature complete.
From the draft specification you will see there is an expire_timeout argument of the org.freedesktop.Notifications.Notify method. That should fit your time out requirement, although I've not used it personally. There is also a org.freedesktop.Notifications.NotificationClosed signal that will allow your program to be notified when the notification is closed, including because of a time out or if it was dismissed by the user.
This post about the screen lock re-design for GNOME Shell 3.10 might give some indication of what notifications are capable of. The post includes some screenshots of notifications appearing in the lock screen.
I'm pretty sure the answer to this is "no" but I figured I'd share the question anyway in case others have a clever workaround :)
I'm building a recipe action so the user could say "Let's make tortilla soup" and then say "next step" to move on to the next part of the recipe. Between each step there might be a long pause as the user is cutting vegetables, etc. Is it possible to have Home either indefinitely wait for a user response or wait for several minutes? Currently it'll wait a few seconds and say something like "Sorry I didn't understand that response" and eventually quit the action.
Forcing the user to go through the "OK, Google, let me talk to the Chef" action over and over is pretty annoying and, I assume, would require immediately ending the conversation after every step, otherwise the device will hang and say "Sorry I didn't understand".
Update
I've found a pretty hacky way of doing this by abusing SSML. There are two options, you can stack up <break/>s
assistant.ask(`<speak><break time="120s"/><break time="120s"/></speak>`);
This actually causes the Home to play a really weird droning noise. Something the devs might wanna look at :D
Another option (which avoids the drone) is to play a 2 minute silent audio clip. According to the docs, 2 minutes is the limit for <audio> but you can stack them up. I just verified that you can make it sit there for at least ten minutes.
assistant.ask(`<speak><audio src="https://.../pause.mp3">hello</audio><audio src="https://.../pause.mp3"></audio></speak>`)
You cannot speak directly to the agent while its paused like this but you can say "OK Google, [whatever command]" and that command will actually get passed to the agent.
There's not currently a way to have the Assistant wait indefinitely for a response.
While technically possible, an app using the suggested workaround to delay responding wouldn't pass the review process - the policies document mentions avoiding playing a silent sound file and communicating for a period in excess of 120 seconds.
A different approach could be to have your app remember the current position in a recipe, but end the conversation after each step. You could then use action invocation to allow the user to say something like "OK Google, ask [your app name] to continue the recipe", jumping back into the conversation and hearing the next step.
There is not, but there is a recently announced feature that might help you do what you're trying.
When answering, you can give the reply and play an audio file using the Media Control. This has several advantages over using the SSML approach you give:
You'll get an event when the audio finishes, so you can prompt again with the guide, or a tip, or a reminder that your action is still there... and then play more audio while you wait.
At any time, the user can say "Hey Google, next step" and your server will get the message and you can send the next step. Or they can say "hey Google, repeat that" and you'll get that message and can repeat the instructions.
They can also ask other questions of your Action that you can answer.
I've decided to integrate OpenFeint into my new game to have achievements and leaderboards.
The game is dynamic and I would like user to be rewarded immediately for some successful results, but as it seems for me, OpenFeint's achievements are a bit sluggish and it shows visual notification only when it receives confirmation from the server.
Is it possible to change something in settings or hack it a little bit to show notification immediately as soon as it checks only local database if the achievement has not been unlocked it?
Not sure if this relates to the Android version of the SDK (which seems even slower), but we couldn't figure out how to make it faster. It was so unacceptably slow that we started developing our own framework that fixes most of open feint's shortcomings and then some. Check out Swarm, it might fit your needs better.
There are several things you can do to more tightly control the timing of these notifications. I'll explain one approach and you can use this as a starting point to explore further on your own. These suggestions apply specifically to iOS apps. One caveat is that these suggestions refer to internal APIs in OFSDK 2.8 for iOS and not ordinarily recommended for high level use and subject to change in future versions.
The first thing I recommend is that you build the sample app with your own product key. Use the standard sample app to experiment before applying the result to your own code.
You are going to get the snappiest response by separating the notification pop-up UI from the process of submitting the achievement. This way you don't have to worry about getting wrapped up in the logic for deciding whether the submission is going just to the local db or is doing the full confirmation on an async network transaction.
See the declaration of "showAchievementNotice" in "OFNotification.h". Performing a search in the sample app, you will see that this is the internal API used for displaying the achievement pop-up when an achievement is earned. It does not actually submit the achievement. You can call this method directly as it is called from "OFAchievementService.mm" to directly control when the message appears. You can then use the following article to disable the pop-up from being called when the actual submission occurs:
http://support.openfeint.com/dev/notification-pop-ups-in-ios/
This gives you complete freedom to call the submission at a later time provided you keep track of the need to do so. For example, you could locally serialize a flag to take care of the actual submission either after the level is done or the next time the app starts up. Don't forget that the user could quit out of a game without cleanly finishing a level.