Learn about FERN-based matching

FERN is Reltio's LLM-based technology for matching. It enhances traditional entity resolution by leveraging advanced pretrained knowledge. As an LLM, FERN applies various kinds of knowledge it has absorbed about the world (language semantics, geography, sociological patterns, name statistics, and colloquialisms) to the goal of matching entities.

FERN is available today for the Individual entity type and the Organization entity type, FERN should not be considered a replacement for manually configured match rules. Instead, think of FERN as an additional match engine, identifying matches in parallel with your configured match rules. For more details, see topics Individual entity type for B2C Data Domains and Organization entity type for B2B Data Domains. It's useful to consider that as an LLM, FERN is able to detect similarities across attributes and records that aren't possible with manually-configured rules. Conversely, you may have some matching requirements that are so explicit, that manually-crafted rules are the most efficient approach. That is why an optimized matching solution may very well involve a blend of both approaches.

For more information about which entity types in each velocity pack support pretrained FERN models and which model applies, see FERN matching support.

Similar to the behavior of manually crafted rules, FERN can produce potential matches and auto-merges. From the Reltio Console, you are able to declare what range of FERN scores should result in potential matches and what range should result in auto-merging. In this way, potential matches can come from both your manually-configured rules and/or via FERN scores.

FERN normalizes relevance scores from deterministic and AI-driven matching to provide consistent scoring across matching methods. This consistency makes match relevance scores easier to interpret when you configure relevance score ranges for potential matches and automatic merges.

Naturally, you will begin by declaring all ranges of FERN scores to result in potential matches. Then, as you gain comfort through analysis. you'll likely promote FERN's higher-range scores to an auto-merging effect In the manner then, auto-merges will likely come in part from your manually-configured rules and in part from FERN.

By analyzing a good sample set of these scenarios, you'll be able to gauge the quality and reliability of matches found by your rules versus those found by FERN, enabling you to determine an optimized blend of configured match rules and FERN-based matching to meet your matching goals, and ultimately develop a strategy that empowers auto-merging.

For more information, see topic Enable a pretrained FERN model.