# Automated Embedding (autoEmbed) quantization types > Source-anchored research for Automated Embedding (autoEmbed) quantization types. Run status: COMPLETED. Parent: [Atlas Vector Search automatic quantization and rescoring](https://llms-explorer.com/tree/atlas-vector-search-automatic-quantization-and-rescoring/) · 5 facets · 14 facts · page: https://llms-explorer.com/tree/automated-embedding-autoembed-quantization-types/ ## Structure and components - autoEmbed float stores each embedding component in float32 form and defaults to dotProduct similarity; MongoDB describes it as the accuracy-oriented option with the greatest RAM and storage requirements among these choices. (high confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - At the component level, converting float32 to eight bits gives a fourfold reduction and converting float32 to one bit gives a thirty-twofold reduction; these arithmetic ratios exclude retained originals, index structures, and other overhead. (high confidence) — [source](https://huggingface.co/blog/embedding-quantization) ## How it works - Mongoid auto_embed_field passes a supplied quantization into the index definition and omits that property when its argument is nil. (low confidence) (verified as of 2026-10-24) — [source](https://github.com/mongodb/mongoid/blob/master/lib/mongoid/fields.rb) ## Problems, failure modes and limitations - Migronaut documents autoEmbed quantization changes as conflicts that cannot change in place. Its prescribed migration creates a differently named index, waits for readiness, moves queries, then removes the old index. (medium confidence) (verified as of 2026-10-24) — [source](https://migronaut.vercel.app/guide/troubleshooting) ## Comparisons and alternatives - autoEmbed scalar uses int8 embeddings supplied by the model, defaults to cosine similarity, and is documented to reduce embedding storage and RAM to roughly one quarter of float. This estimate concerns embeddings rather than an entire deployment. (high confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - The driver tracking issue identifies quantization and numDimensions as additions requiring index-builder/type support; check the installed driver rather than assuming all historical builders expose them. (medium confidence) (verified as of 2026-10-24) — [source](https://jira.mongodb.org/browse/DRIVERS-3389) ## Facts and statements - autoEmbed accepts float, scalar, binary and binaryNoRescore; omitting quantization selects scalar. (high confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - float stores 32-bit values and defaults to dotProduct similarity; scalar uses model-provided int8 values and defaults to cosine. (low confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - binary stores one-bit index values plus full-precision counterparts for candidate rescoring; MongoDB estimates RAM at 1/24 of float and defaults to euclidean similarity. (low confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - Do not copy the quantization default from a precomputed-vector field into an autoEmbed definition: vector fields use none as the automatic-quantization default, whereas autoEmbed uses scalar and names its full-precision option float. (high confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - autoEmbed binary derives one-bit index values from full-precision model output, retains the full-precision vectors, and rescores selected leading candidates with those vectors; its default similarity is euclidean. (high confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - MongoDB estimates autoEmbed binary RAM at approximately 1/24 of float. Its retained full-precision vectors mean this RAM estimate is not an equivalent guarantee of disk savings. (high confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/vector-search/indexes/vector-search-type.md) - Benchmark quantization on representative embeddings and queries before choosing it: independent embedding experiments and vector-engine documentation show that quality depends on model and data, and that rescoring can recover quality while adding retrieval work. Those results do not establish a MongoDB-specific recall guarantee. (high confidence) — [source](https://huggingface.co/blog/embedding-quantization) - Treat a blanket claim that binary quantization always preserves accuracy and accelerates queries as unsupported: Qdrant scopes one-bit effectiveness to vector characteristics and warns that disk-backed rescoring can reduce search speed; this is external caution, not an Atlas benchmark. (low confidence) — [source](https://qdrant.tech/documentation/guides/quantization/)