We're excited to announce that the default model for new projects in Voiceflow has been updated to Claude 3 - Haiku. This change reflects our commitment to providing the most advanced and effective tools for creating AI agents.
What’s New:
Claude 3 - Haiku is now the default model in all new Voiceflow projects, offering enhanced performance, improved language understanding, and more nuanced conversational abilities.
This update does not affect existing projects, but you can manually switch to Claude 3 - Haiku within your project settings if you wish to take advantage of the latest models.
As part of our ongoing efforts to improve the functionality of Voiceflow, we are introducing an update to the Set step data format: the transition from setV2 to set-v3 nodes. This update provides a more refined structure to the Set step, offering improved control and flexibility for setting variables in your projects.
Key Changes
Transition to set-v3 Nodes:
We are transforming the current setV2 nodes into set-v3 nodes. This migration will ensure a more robust and flexible approach to variable management, with an updated structure that simplifies the process of setting variables within your projects.
Automatic Conversion:
All existing setV2 nodes will be automatically converted into set-v3 nodes when a project is opened in-app. This conversion preserves all existing functionalities while enabling you to leverage the new and improved system.
Technical Overview
The new set-v3 structure is outlined below:
Properties Affected:
Fields Added: None
Fields Modified:diagram.nodes[type=setV2] transformed into diagram.nodes[type=set-v3]
Fields Removed: None
Migration Example
To illustrate how the migration transforms setV2 nodes into set-v3 nodes, here are examples before and after the migration:
As we approach the release date for this migration, we encourage all users to familiarize themselves with these changes to make the most of the enhanced Set step functionality. If you have any questions or need assistance, our support team is ready to help. Stay tuned for the release!
Executing “untrusted” code is tricky. Bad actors can write malicious code and potentially access sensitive data.
The current way the javascript step is set up is a security risk and we want to move off of it at the earliest opportunity. Luckily there is a new backend that is both secure and more performant:
It's now 70-94% faster, there will be more information about that in another post.
All Javascript steps created after July 30th, 2024 already automatically use the new system. We’ll be slowly converting existing Javascript step to use the new system, with the cutoff by August 16, 2024.
The goal of the javascript step is to provide quick scripting to manipulate variables, rather than be a heavy-load serverless function with networking. For that, we can use functions.
No action is needed is on your end, unless you use the following patterns.
We've monitored all javascript step errors for the past week, running both the new and old backends in parallel, to categorize impact and effects.
For the select few users that are affected, we will be proactively reaching out to them.
All breaking changes we've observed will all be documented here, so people don't reimplement in the future.
Major breaking changes
requireFromUrl
The javascript step used to support requireFromUrl() which allows users to load in 3rd party libraries via URL. Commonly libraries such as moment or lodash and other utilities. This method is actually an anti pattern and has major security risks that the new backend does not support.
// example usage of requireFromUrl
const moment = requireFromUrl("https://cdnjs.cloudflare.com/ajax/libs/moment.js/2.29.1/moment.min.js");
time = moment().add(1, 'hours');
This will no longer work after the cutoff, the Javascript step will go down the fail port with a debug message saying: ReferenceError: fetch is not defined"
nodeJS modules: buffer
Our new backend does not run on NodeJS, but rather a v8 isolate.
Calls specific to NodeJS modules rather than the Javascript/ECMAScript standard will no longer be supported.
// example of using nodeJS modules
let buff = new Buffer(token, 'base64');
name = buff.toString('ascii');
This will no longer work after the cutoff, the Javascript step will go down the fail port with a debug message saying "ReferenceError: [module] is not defined".
There are low level alternatives to replicate the behavior of nodeJS utilities, and this is something LLMs excel at helping convert (be sure to test, of course!).
If you are using Buffer for base64 encoding/decoding, you can easily polyfill a atob or btoa function:
function atob(e){let r="";if((e=e.replace(/=+$/,"")).length%4==1)throw Error("Invalid base64 string");for(let t=0,n,a,o=0;a=e.charAt(o++);~a&&(n=t%4?64*n+a:a,t++%4)&&(r+=String.fromCharCode(255&n>>(-2*t&6))))a="ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/".indexOf(a);return r}
function btoa(t){let a="ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/",h="",r=0;for(;r<t.length;){let e=t.charCodeAt(r++),$=r<t.length?t.charCodeAt(r++):Number.NaN,c=r<t.length?t.charCodeAt(r++):Number.NaN;h+=a.charAt(e>>2&63),h+=a.charAt((e<<4|$>>4&15)&63),h+=isNaN($)?"=":a.charAt(($<<2|c>>6&3)&63),h+=isNaN(c)?"=":a.charAt(63&c)}return h}
// converts base64 to ascii, equivalent of Buffer.from(base64ref, 'base64')
atob(base64ref)
// converts ascii to base64
btoa("hello world")
JSON.stringify(this)
The this keyword now has a circular reference, and calling JSON.stringify(this) will result in recursion:
TypeError: Converting circular structure to JSON
--> starting at object with constructor 'global'
--- property 'global' closes the circle"
You can still use this.
Other minor changes
DateTimeFormat
Intl.DateTimeFormat() constructor. As per the MDN documentation:
If any of the date-time component options is specified, then dateStyle and timeStyle must be undefined.
so Date(A).toLocaleTimeString.(B, { timeZone, timeZoneName, timeStyle }) would crash.
fetch
The fetch() command already doesn't work today - we don't really support any await or async actions, but after the cutoff, the Javascript step will go down the fail port, with a debug message saying "ReferenceError: fetch is not defined".
resource changes
The new backend solution has the same timeout but will have a definited limit on CPU and memory.
In our monitoring no present blocks are affected by these limits.
As AI agents continue to evolve, we have observed a significant shift towards building on Web chat and deploying on channels like WhatsApp, Slack, and Discord using our Dialog Manager API. To better align with this trend and allocate our resources effectively, we have decided to discontinue our direct hosting integration with Alexa.
Impact
Existing Alexa skills managed and deployed on Voiceflow will no longer function.
We strongly encourage you to migrate your skills directly to Alexa, as we will not be providing updates or support for our Alexa integration.
Support
We understand this transition may be challenging for some of our customers. Our team is here to support you during this period. If you have any questions or concerns, please reach out to us. Additionally, we have a growing Discord community with tutorials and community members ready to assist you.
Thank you for your understanding and continued support. We look forward to working with you to build amazing AI agents on our enhanced platform.
We are excited to announce the addition of GPT-4o mini to our platform. This integration brings new capabilities and improvements to enhance your AI agent experience. Below are the details of this update:
New Features:
GPT-4o mini Integration: Our platform now supports GPT-4o mini, providing exceptional cost-efficiency and speed. This compact model is perfect for many of your agent tasks.
Improvements:
Cost Efficiency: Enjoy significant cost savings with GPT-4o mini with a 0.2x token multiplier.
Faster Performance: Experience quicker response times, ensuring your AI agents perform efficiently.
We are committed to continuously improving our platform and providing you with the best tools available. Thank you for your continued support!
We are announcing the sunset of the FAQ API, effective August 6th, 2024. This decision is based on user feedback and our commitment to providing the best tools for your conversational AI needs. We believe that our new feature for uploading tabular data to the Knowledge Base is a more efficient and versatile solution for managing FAQs.
To help you transition smoothly, we have created a short video demonstrating how to use this new feature. You can watch the video below.
We appreciate your understanding and cooperation as we make this transition. Please reach out to our support team if you have any questions or need assistance.
We're proud to announce Voiceflow's newly revamped documentation. You can find it at docs.voiceflow.com.
This upgrade seeks to streamline all learning resources to help you use Voiceflow's designer tools as well as the APIs. Below are the details of this update:
Unified documentation: The old learn.voiceflow.com and developer.voiceflow.com sites have had all their articles merged into our new domain, docs.voiceflow.com. This is now the only site you'll need to visit to access all of Voiceflow's documentation. No more separate designer and developer documentation splits. Using the old domains will still redirect you to docs.voiceflow.com. Links inside the Voiceflow platform have also been updated.
New organization: We've reorganized our documentation into Building, Deploying, and Improving Agents, as well as new docs and guides for Getting Started and Improving agents.
Updated documentation: Many new articles have been written to teach you more about the state of the modern Voiceflow platform.
New API docs and guides: Our API docs have been restructured, with new conceptual articles added and a guide for Getting Started with APIs.
Moved changelogs: We've also brought over our changelogs from the old website (changelog.voiceflow.com). The link will still work to reach the changelogs.
You can learn more about our new docs in this video.
We are committed to continuously improving our resources and providing you with the best documentation available. Thank you for your continued support!
Effective August 12th, we will be sunsetting our native integrations for WhatsApp, Twilio SMS, and Microsoft Teams. The project types for these channels will remain accessible, but all active connections to these channels will be severed, and it will no longer be possible to publish new projects or maintain existing ones on these platforms.
We understand this change may impact your workflows, and we are here to support you through this transition. Please explore our other integration options or contact our support team for assistance in adapting your projects.
Thank you for your understanding and continued support.
We are excited to announce the addition of Gemini 1.5 Pro, the first Google model we support, to our platform. This integration brings new capabilities and improvements to enhance your experience. Below are the details of this update:
New Features:
Gemini 1.5 Pro Integration: Our platform now supports Gemini Pro 1.5, offering enhanced performance and a broader range of functionalities. This model brings advanced natural language understanding and generation, making it ideal for various applications.
Improvements:
Enhanced Accuracy: With Gemini 1.5 Pro, you can expect improved accuracy in natural language processing tasks, leading to more precise and reliable outputs.
Faster Response Times: Enjoy quicker response times, thanks to the optimized performance of Gemini 1.5 Pro.
Broader Language Support: Gemini 1.5 Pro offers support for more languages, providing a better experience for multilingual users.
We are committed to continuously improving our platform and providing you with the best tools available. Thank you for your continued support!
We are thrilled to introduce new management tools that enhance your ability to oversee and scale your AI agents. Our latest update offers improved visibility into the connections between your agent data and project workflows and components, making it easier to understand, manage, and grow your AI projects effectively.
What’s New:
Intents, Components, and Functions tables
A new “Used By” column has been added to these tables.
This column displays which workflows or components are utilizing each resource.
Each item in the dropdown is a direct link.
Easily navigate to the referenced workflow or component for quick access and management.
Benefits:
Improved Visibility
Understand how all parts of your agent design come together at a glance.
Quickly identify dependencies and interconnections between resources.
