# Atlas Vector Search RAM sizing and automatic quantization > Source-anchored research for Atlas Vector Search RAM sizing and automatic quantization. Run status: COMPLETED. Parent: [Platform Consolidation TCO (Atlas Search/Vector Search vs Separate Search or Vector DB)](https://llms-explorer.com/tree/platform-consolidation-tco-atlas-search-vector-search-vs-separate-search-or-vector-db/) · 5 facets · 12 facts · page: https://llms-explorer.com/tree/atlas-vector-search-ram-sizing-and-automatic-quantization/ ## How it works - The Atlas index-list Required Memory metric approximates query-time memory needs; Size measures total index storage and includes full-fidelity disk vectors. Use Required Memory for the resident-index budget and allow additional process memory. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/vector-search/about/vector-quantization.md) - TCO inference: do not apply a published vector-memory reduction directly to the total Atlas bill. Full-fidelity disk storage remains, process memory is also required, and HighLevel found memory relief insufficient to resolve its wider workload constraints. (medium confidence) — [source](https://www.mongodb.com/docs/vector-search/deployment/deployment-options.md) ## Measurements and reference values - MongoDB documents approximate vector-index RAM reductions of 3.75x for scalar and 24x for binary quantization. The vector values shrink 4x and 32x, respectively; the HNSW graph remains uncompressed. These are not whole-deployment cost-reduction factors. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/vector-search/deployment/deployment-options.md) - Unquantized vector capacity scales with vector count and dimensions. MongoDB gives roughly 3 KB per 768-dimensional vector, before additional metadata and process headroom; collection storage bytes are not a substitute for vector-index memory. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/vector-search/deployment/deployment-options.md) - TitanBases reports scalar quantization reduced index memory from 4.51 GiB to 1.36 GiB in a controlled accuracy test associated with its eKYC POC. Recall@10 fell from 99.37% to 92.83%, and Exact Match@10 from 94.88% to 41.22%; the team rejected that loss for identity deduplication. These figures are workload-specific. (low confidence) (verified as of 2026-10-02) — [source](https://titanbases.com/blog/mongodb-atlas-vector-search-ekyc-deduplication) ## Problems, failure modes and limitations - HighLevel reports that quantization materially reduced memory and latency on a roughly 48,000-document knowledge base. Scalar preserved quality better than binary on its slow-query set. The team treated scalar as temporary relief and still migrated, showing that RAM savings alone did not resolve that deployment's scaling problems. (low confidence) (verified as of 2026-05-04) — [source](https://engineering.gohighlevel.com/post/ai-knowledgebase-vector-db-migration) ## Comparisons and alternatives - For supplied embeddings indexed with type vector, fields.quantization defaults to none; select scalar or binary explicitly. Automatic means Atlas performs conversion at indexing and query time. Changing the setting rebuilds the index. This scope differs from the autoEmbed field type, whose documented default is scalar. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/vector-search/about/vector-quantization.md) - MongoDB recommends approximate disk-to-RAM capacity ratios of 4:1 for scalar and 24:1 for binary quantization, and free disk space equal to 125% of estimated index size. These are sizing guidelines rather than guaranteed utilization ratios. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/vector-search/deployment/deployment-options.md) ## Facts and statements - Automatic quantization keeps compressed vectors in memory and retains full-fidelity vectors on disk for exact search and binary rescoring. Total persisted index size can therefore increase while required RAM decreases. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/vector-search/about/vector-quantization.md) - MongoDB warns that automatic binary quantization can shift the binding capacity constraint from RAM to storage. For TCO estimates, smaller resident vectors alone do not establish that a smaller instance can hold the persisted index. (medium confidence) (verified as of 2025-08-12) — [source](https://www.mongodb.com/docs/vector-search/deployment/deployment-options.md) - Independent operator evidence supports measuring application quality before counting quantization savings in TCO: TitanBases rejected scalar for identity matching, while HighLevel found scalar useful but insufficient for its broader workload. Neither report establishes a universal Atlas quantization accuracy penalty. (medium confidence) — [source](https://titanbases.com/blog/mongodb-atlas-vector-search-ekyc-deduplication) - A MongoDB product-team update dated April 29, 2025 announced general availability of automatic scalar quantization, binary quantization with rescoring, and int1 ingestion, together with approximate Required Memory reporting in the Atlas UI. (low confidence) — [source](https://www.mongodb.com/community/forums/t/automatic-scalar-and-binary-quantization-now-supported/306377)