SABLE is the filter-native retrieval engine driving VectorAmp: a large-scale vector index that treats the WHERE clause as a first-class part of the query plan. In the whitepaper's W1/W2 single-category and IN-list filter workloads at 100 million vectors and 10 percent selectivity, SABLE was more than 28× faster than DiskANN across the reported cases. One reported case measured 47 milliseconds median latency against 2.9 seconds on the same node, more than 60× faster, at Recall@10 of 0.96 or better.
Your permission model is only as good as the layer that enforces it. If filtering is something your database does after retrieval, late filtering can waste work and return too few eligible results. SABLE makes the requested scope part of the query plan so teams can evaluate speed, recall, and permissions together.
Single-node results on stock AWS instances. Full methodology, query workloads, and ablations live in the whitepaper.
*More than 60 times smaller architectural index-memory target than conventional FP32 HNSW with M=16 at comparable recall. SABLE keeps compressed search structures and full-precision reranking vectors on NVMe. This compares index structures, not whole-service memory or query latency.
Four deliberate design choices that compound at scale.
Most vector stores treat metadata filters as a post-processing step that throws away precomputed work. SABLE uses metadata predicates to narrow the candidate space before graph-free ADC search, with shard-local statistics guiding the work.
Sharding is part of how SABLE plans and routes queries. The query coordinator routes by metadata keys, prunes shards by filter, and merges candidates for exact reranking. Capacity is designed to expand with additional shards.
SABLE is tuned for the retrieval patterns VectorAmp itself runs on: filtered semantic search, RAG with citations, recommendation, and deduplication. Practical defaults for product workloads, not synthetic-benchmark theater.
Hot paths should stay fast, warm data should be compact, and cold data should not dominate your bill. SABLE is designed around explicit storage tiers and predictable read amplification.
SABLE is a compact cold-tier vector index with filter-native planning built into the storage layout, not bolted on after retrieval.
Python and TypeScript clients use the same dataset-scoped search API as the platform. Pass text or vectors, top-k, metadata filters, and advanced range filters directly.
# filtered semantic searchfrom vectoramp import VectorAmp
client = VectorAmp(api_key="va_...")dataset = client.datasets.get("dataset_id")
results = dataset.search( text="migration plan risks", top_k=20, filters={"team": "platform", "region": "eu"}, advanced_filters=[{"field": "year", "op": "gte", "value": 2025}], include_documents=True,)// same query, TypeScriptimport { VectorAmp } from "@vectoramp/vectoramp";
const client = new VectorAmp({ apiKey: process.env.VECTORAMP_API_KEY });const dataset = await client.datasets.get("dataset_id");
const results = await dataset.search({ queryText: "migration plan risks", topK: 20, includeMetadata: true, filter: { team: "platform", region: "eu" }});Historical single-node whitepaper results on 1M-vector DBpedia-OpenAI and 100M-vector Deep-1B workloads. Latency is labeled per row. These are workload-specific comparisons, not a production SLA or a claim about other datasets.
| Filter-nativeSABLE | Disk-resident graphDiskANN | In-memory graphHNSW | |
|---|---|---|---|
| Unfiltered p95 (1M, R@10≈0.94) | 5.1 msnp=16, rerank=4 | 175 msdisk-resident | Fast in-RAMat huge memory cost |
| Filtered p95 (1M, ~11% selectivity) | 14–20 mspredicate pre-gates lists | 132.9 msfilter as post-step | Degrades sharplyfilter masks during traversal |
| Filtered p50 (W1/W2, 100M, 10% selectivity) | 24–59 ms28–101× faster · single-category + IN-list filters | 1.3–3.1 sdisk seeks dominate | Memory-boundconventional FP32 HNSW |
| Memory @ 1M, d=1536 | ~96 MBarchitectural index-memory target | Disk-resident | ~6.27 GBFP32 HNSW, M=16 |
SABLE supports shard routing and configurable search parameters. Size and validate your own filtered workload with our team.
Managed SABLE Cloud is available. Customer-operated BYOC is being scoped as an early-access pilot, subject to installation and support review. The initial scope is the core SABLE stack; ingestion, embedding, reranking, and Intelligence are future extensibility, not included pilot promises.
Fully managed on AWS. The same SABLE engine that powers the VectorAmp Intelligence Platform, surfaced through the VectorAmp API and SDKs. A hard-capped trial is available without a credit card.
For teams evaluating larger workloads or customer-operated deployment. BYOC early access is a scoped pilot, not a generally available install. CPU training and optional GPU training are evaluated by profile; supported hardware and operating responsibilities are confirmed during scoping.
Yes. The same engine. The VectorAmp platform runs on SABLE. Buying SABLE standalone gets you the same engine, the same indexing pipeline, and the same query planner without the application layer (canvases, agents, intelligence UI). If you outgrow SABLE-only and want the platform on top, your indexes carry over.
In SABLE, filtering is part of the index, not a separate step layered onto vector search. Each IVF list carries categorical and numeric summaries so SABLE can eliminate lists that cannot satisfy the filter, narrow the candidate set, run graph-free ADC over compressed vectors, and then rerank against full-precision vectors. That planning is what we mean by filter-native.
SABLE is the index VectorAmp creates for datasets: an OPQ-rotated IVF-PQ foundation, graph-free ADC search, exact reranking, and per-list metadata used to narrow filtered work. We benchmark against systems like HNSW and DiskANN, but we do not expose those as customer-selectable index modes in the current SDKs.
SABLE is not a relational database and does not provide joins. Current filter-native planning focuses on equality and numeric-range predicates. Recall is high but workload-dependent, not 1.0, and the scoped BYOC pilot is not an air-gapped deployment.
Managed VectorAmp Cloud is available. Customer-operated BYOC is being scoped as an early-access pilot, subject to clean-account installation and support evidence. Contact sales to explore a scoped pilot.
SABLE is included in the VectorAmp Cloud plans. The trial is free with hard-capped usage; paid plans include monthly vector-query and vector-write allowances with metered usage above them. See the pricing page for the current plan lineup, or contact sales for Enterprise terms.
Stop paying the post-filter tax. Run your filtered workload against the same engine that powers the VectorAmp Intelligence Platform.