Upgini Generative AI can automatically enrich any text fields with relevant facts from external data sources and generate ready-to-use numeric features from enriched representations of text fields.

Upgini Gen AI does this in the following steps:

(1) Finds entities in the text to match facts from external data sources. Simple examples are the company name, car model, product title, and place of interest (POI) name.

(2) Detects contextual information for extracted entities from external sources. A simple example is a geographical location for a company.

(3) Generates enriched embeddings for text fields using a process similar to Retrieval-augmented generation (RAG) with facts fetched from external sources and specific contextual information.

These enriched embeddings for text fields can be used as numeric features to enhance the accuracy of downstream ML models.

We want to improve the accuracy of an ML model that predicts the probability of product usage decline for a specific client (what is called an attrition or churn model).

For every client in a labeled training dataset, in addition to numeric ML features, we have transcriptions of all calls to support, the history of the client's chats with support, and purchased product reviews from the website.

Simply pass this text as columns in a labeled dataset during the search process, and Upgini will automatically enrich these columns with relevant external sources and generate Enriched embeddings.

Enriched embeddings from Upgini Generative AI are more accurate than embeddings from top commercial embedders, such as OpenAI’s ada-002.

More details on the comparison can be found in the Medium publication.