How Filtered's AI handles your data

Every generative AI feature in Filtered (such as our AI-assisted search and our AI-assisted pathway curation feature) used to run on an integration with OpenAI. We have replaced that with a self-hosted large language model running entirely within our own AWS environment. Whenever a message is sent through an AI feature, whether that is the chat companion today or buttons and other interactions in future, the information now stays inside the Filtered environment.

Our previous setup used an OpenAI enterprise account. Under that arrangement, data sent through the API was covered by contractual rules and was not used to train OpenAI's models, which is different from typing into the consumer version of ChatGPT. So your information was never used for training. Even so, we strictly limited the context sent to OpenAI to minimise the data transfer, which in turn limited the capabilities of the genAI features.

Now, with AI hosted in-house:

🔒  Why this is more secure. Previously there was no way to use OpenAI without sending data to servers based in the US, which in some cases required international data transfer agreements under global and UK data protection regulations. Keeping everything on our own infrastructure within the EU removes the need for such international agreements, and it means we no longer need additional any third-party data handling agreements to support these features. All data shared with Filtered stays within servers in Ireland or Germany.

Because the data is no longer leaving Filtered, we can give the AI far more context about your instance. That turns it into a genuine partner in the product rather than something we had to keep at arm's length.

Today the search companion uses the chat to form a search term, and that term is then run inside Filtered. The chat itself has no context about the assets in your instance. With AI in-house, the companion can access Filtered search directly and read metadata about your assets, with no risk of data leaving the environment. That makes it far more capable: better recommendations, better content discovery, and the ability to draw on transcripts from the files we process. For example, the AI can now tailor the search to particular types of resources, or recent resources, with ease.

It is now much easier for us to add small AI-powered touches throughout the product, because we can send short queries to our own LLM and pass it context about your business and about the person using it. An example (that we have wanted to build for some time) is a “define this skill” button on the skills page, which we can now do without sending anything to an outside provider. You can expect AI-boosted features everywhere in the product now, safe in the knowledge that no data is being passed to OpenAI or another big AI provider.

Because we can now safely expose much more of what is in your instance, the AI can act as a genuine co-pilot across the platform. That includes producing content value analyses, defining skills, building skill frameworks, and describing your content. Automated pathway building is one of the bigger things this opens up, so we will be able to move much more pathway creation towards automation. You’ll find you can do a lot more with Filtered to support your enterprise skills and AI strategy - without burning through tokens for Claude or ChatGPT!

ℹ️  A note on the changes: The move to in-house hosting is complete. The richer capabilities described above, such as direct search access, the ‘define this skill’ button, and automated pathway building, will roll out over the next few months.

Do you have questions? Contact success@filtered.com