MongoDB has released two reranker models designed to make AI applications better at retrieving data for users.

An LLM finds the information needed to answer a prompt in two steps. First, it scans the data sources at its disposal and identifies files that might contain the necessary information. It then ranks those files based on their relevance to the query. The latter task is performed with AI models dubbed rerankers.

MongoDB’s two new rerankers, rerank-2.5 and rerank-2.5-lite, were developed by its Voyage AI subsidiary. Their context length is twice as large as that of their predecessor, which means they can ingest more data per query. They also process that data with higher accuracy. The models scored more than 7% better than Cohere Rerank v3.5, a competing algorithm, in an evaluation that measured how well they rank files by relevance.

MongoDB has equipped rerank-2.5 and rerank-2.5-lite with what it describes as an instruction-following capability. The feature allows users to manually customize how the rerankers should prioritize files. For example, an accountant could instruct the algorithms to only sift through spreadsheets when answering a question about next quarter’s earnings.

Photo courtesy of MongoDB