Most production RAG systems are built on a simple premise: convert documents into single vectors and find the ones closest to a query vector. This works for simple documents, but fails on the messy, multi-aspect data that defines enterprise reality. Cohere's Compass is a new embedding model designed for this specific problem, and it suggests a necessary evolution in how we build retrieval systems.
the single-vector problem
Standard embedding models, including powerful ones like Cohere's own Embed v3, map an entire document to a single point in semantic space. This is a lossy compression. If a document contains multiple distinct concepts—like an invoice with a specific sender, due date, and line items—the resulting vector is an average of all those concepts. The relationships between them are lost.
This leads to retrieval errors that are painfully familiar to anyone who has shipped a RAG product. A search for a "red T-shirt" might return "blue and yellow jeans" because the vector for colors is muddled with the vector for clothing type. In an enterprise context, a query for an invoice from a specific person might fail because the contextual link between the sender and the attached document was severed during the chunking and embedding process.
