Filtered uses two distinct AI systems that work together. The first is a core machine learning algorithm, built and maintained entirely by Filtered, which handles natural language search and indexing. The second is an internally hosted large language model (LLM), running on Filtered’s own secure servers, which powers the platform’s agentic and generative features. Anything in the platform that depends on understanding content, or on understanding skills, draws on one or both of these systems.

The core algorithm is a machine learning model for natural language search and indexing. It is hosted, maintained and optimised by Filtered, and it works across all major languages. Every search result in the platform is produced by this algorithm, which means it sits behind:

At its simplest, the algorithm assesses how relevant each piece of learning content is to the skills that matter for a given client. That assessment of relevance can be made against a skill, a search term or a playlist brief.

Asset comprehension. Each client’s content catalogue is indexed into what we call a ‘meaning space’, based on the descriptive text that comes with the content, such as titles and descriptions. The algorithm converts this natural language text into a numerical representation using a sentence transformer. The transformer relies on a ‘self-attention’ mechanism, which works out which parts of the descriptive text matter most to its meaning and weights them accordingly.

Once converted, every piece of content is positioned as a point in a high-dimensional meaning space, where points that sit close together represent content covering similar material. Filtered’s production database holds over 14 million pieces of learning content, so this is a computationally intensive step. It runs in the background, which is why content can be prioritised and filtered for any newly required skill within seconds.

Skill representation. When a new skill, playlist or search term is defined, the algorithm follows the same process using the skill description. It produces a numerical representation of that skill and compares it against every piece of content already mapped in the meaning space. A relevance score is calculated for each piece of content, and that score is what ranks and filters your results.

We benchmark the algorithm against human experts. In a controlled experiment with a global pharmaceutical client, using their own skills taxonomy and real assets from their LMS, the algorithm reached an overall precision of 83%.

To put that figure in context, precision measures how ‘pure’ the tagged output is: the proportion of tagged assets that are tagged correctly. We use skilled human curators as the gold standard for what ‘correct’ looks like. When we compare one expert’s tagging against another’s, agreement typically falls in the 60 to 80% range depending on the skill. Algorithmic tagging is therefore at least as precise as expert human tagging, while working at a scale no team of people could match.

On top of the core algorithm, Filtered runs an internally hosted large language model. This is a frontier model that runs on Filtered’s own secure servers, which keeps client content and metadata inside our controlled environment. The LLM handles the platform’s agentic and generative features:

The core algorithm and the LLM are built to complement each other. The LLM handles interpretation and generation: understanding a file, reading a customer’s intent, or refining a brief. The algorithm handles relevance and indexing at scale.

In practice, the outputs of the LLM are fed into the core algorithm. A query that the LLM has interpreted and augmented is passed to the algorithm, which runs it against the meaning space to produce indexed, ranked results. The LLM works out what to ask, and the algorithm works out which content best answers it.