Nuxeo Insight does not only allow your Nuxeo Applications to be AI-ready but also Enterprise ready. It can add Machine Learning to resolve your use cases while applying it directly with your enterprise production content.
It does it in a simple way that makes it seamless, immediate, scalable, but always in a safe manner so you are always in control.
Insight integration with the Nuxeo Platform provides a feature set that takes care of all data engineer, ML expertise and governance to nurture content bot's evolution.
As a user, you focus is on your domain, your content and your use cases.
As new content is ingested or annotated with metadata, it is automatically exported to your corresponding bots.
This is content is valuable as it brings new examples for the bot to train on. These examples could be edge cases, new categories or even just more information on cases that we need to improve on.
A new train is dispatched to incorporate all this new knowledge on the bot's neural network.
This process keeps the bots up to date and constantly evolving.
When we just add all the new content we have, there is no control over what might be improved. There isn't even a certainty that anything will be improved at all. We could just be adding similar examples to what we already have in abundance. This will not have a great impact on your bots.
Active learning is an intelligent process that selects examples from made predictions that provide a high probability of improving a bot. It puts these predictions in front of a user to be confirmed. By confirming, we can have this critical new content. It will rapidly evolve the bot into great performance.
Nuxeo Insight goes further by providing a way for you to choose what values or fields you need to have better performance, as they are critical for your business. The human in the loop activity will then focus on these values enhancing exactly what are your business needs.
As content bots evolve and start getting confident they will start providing the user with suggestions.
These suggestions are shown on all forms, like the creation or the edit form.
Suggestions are displayed with their confidence level. The user can just tap them to confirm its value on the input.
As each content bot continues to mature a reach a higher confidence level, it will fill in content directly.
Inexistent metadata will be filled, but as human also make mistakes, existing data is also corrected. As some content is mission-critical, you decide the confidence thresholds for each situation.
Content creation is done in bulk on any part of your repository.
Once you've created, trained your content bot and it's ready to work, you need to publish it so it can be used in your Nuxeo Application.
Usually in companies with an ML department, having a made custom Machine Learning model into production is a project breaker for all the needed engineering, I/O contract and even integration between systems.
With Nuxeo Insight, all the orchestration is provided with the click of a toggle. The user activates the bot or a specific version, and it's immediately deployed and available to the instance.
As content bots are build and used, you have a full understanding of how it's performing.
You see each version's performance against new incoming data knowing if you need to retrain it. You also get evaluations with a full test dataset on each new training so you can properly compare them with other versions.
You are provided details like confusion matrix data, each value performance as well as bot usage over time.
On each suggestion, you can know on what content it is based, which version of the bot did it and what was the training example set used to train it.
You can also audit which content was automatically filled in or corrected by a content bot.
This gives you at any point all the needed information to understand Insight actions.
Over time, some edge case situations may happen, where a content bot will see its accuracy reduced with new training. This may occur slowly, but when huge amounts of content are ingested in a go, it may be normal that the current training configuration is not suitable anymore.
These cases might generate wrong content and even mistakenly correct existing content.
At such times, you can backtrack bot actions and make sure that the content is reverted to the previous one. This makes Insight fail-proof as we can always regress actions.