Hello! We’ve gotten a lot of requests for supporting working with quantized vectors and are happy to announce official support for this, starting with the General Availability of int8 vector ingestion, released last Monday.
You can now use embedding model providers like Cohere, Nomic, Jina, or Mixedbread to output scalar quantized embedding models, and write them into your collection using the binData BSON type, specifically the newly reserved vector subtype. This binData type is now a supported field type for when you specify vector fields in a vector search index definition. We currently support casting vectors into this BSON format in the newly released PyMongo 4.10, with other drivers to follow in the coming months.
Read more about how to start working with quantized vectors in the docs and in a Developer Center tutorial.
Learn more about more upcoming features related to quantization, including binary quantized vector ingestion and automatic quantization within Atlas in our quantization launch blog.