Automated Embedding (autoEmbed) quantization types including binaryNoRescore
Parent: Atlas Vector Search automatic quantization and rescoring · Published reference · snapshot 2026-10-24 · skill mongodb-cohort-326a9da785a4
↓ Facts as markdownall context files
Source-anchored research for Automated Embedding (autoEmbed) quantization types including binaryNoRescore. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Comparisons and alternatives
- Independent background evidence: the Hugging Face/mixedbread quantization experiment reports roughly 92.5% retained retrieval performance without binary rescoring versus about 96% with rescoring. It uses float-query/binary-document rescoring, so these figures do not benchmark Atlas binaryNoRescore or its full-precision rescoring alternative. (low confidence) [source]
Facts and statements
- The autoEmbed quantization field accepts float, scalar, binary, and binaryNoRescore; scalar is the default. (high confidence) (verified as of 2026-10-23) [source]
- For autoEmbed, float stores 32-bit values. Scalar uses model-provided int8 values and approximately one quarter of float embedding storage and RAM. (high confidence) (verified as of 2026-10-23) [source]
- For autoEmbed, binary indexes one-bit values but retains full-precision vectors for candidate rescoring. MongoDB states RAM use is about 1/24 of float; this is not a corresponding total disk-storage reduction. (high confidence) (verified as of 2026-10-23) [source]
- binaryNoRescore uses model-provided one-bit vectors without rescoring. MongoDB describes faster queries and lower storage cost than binary, with lower accuracy; it gives no universal latency or recall improvement. (high confidence) (verified as of 2026-10-23) [source]
- MongoDB server documentation lists default similarity as dotProduct for float, cosine for scalar, and euclidean for binary and binaryNoRescore. (high confidence) (verified as of 2026-10-23) [source]
- Mongoid auto_embed_field accepts quantization and passes it into the autoEmbed index field specification. DRIVERS-3389 tracks the builder/type support required to expose these index options through language drivers. (medium confidence) (verified as of 2026-10-24) [source]
- Scope boundary: Percona documents a self-managed autoEmbed pipeline whose Ollama model catalog sets quantization to float. This is evidence of a separate implementation context, not verification of Atlas binaryNoRescore support or performance. (low confidence) (verified as of 2026-10-01) [source]
Children
- No children recorded.