DiskANN / SPANN
Disk-based ANN
For corpora that outgrow RAM: test SSD reads, cold queries, filtered recall and refresh behavior on a real deployment.
Primary reference ↗WORKLOAD LAB / REPRODUCIBLE EXPERIMENTS
Test the workload.
Then choose the index.
Run real search algorithms on shared data. See what compression, filters and search effort do to recall, latency and index size before shortlisting a service.
01
Structural stress tests
Eight seeded workloads with held-out queries, filters and edge cases.
02
Public text retrieval
Real MiniLM embeddings and passage annotations from SQuAD.
03
Deployment evidence
A service runner for concurrency, freshness and decision-ready reports.
Choose a workload
Generated structural fixture
Enterprise knowledge retrieval
How much search effort preserves neighbors after department filters?
1,000
indexed records
100
held-out queries
64
dimensions
10
neighbors requested
Ready to run on this device.
Runs in a background worker. Larger experiments can take a minute. Timings describe this device and these JavaScript implementations.
Dataset provenance & interpretation
How much search effort preserves neighbors after department filters? Independent query random stream; no indexed self-points. Synthetic metadata and texts are not evidence of customer representativeness or semantic retrieval quality. No human relevance labels.
Embedding: Seeded normalized Gaussian/topic construction, 64 coordinates; not a language model
License: Project-authored synthetic fixture; no customer data · Version: structural-v1 · Seed: 4101
Source: Vector Index Lab authored structural fixture
TAKE IT TO YOUR DEPLOYMENT
Measure the service you might buy.
Use the same exported workload against your own endpoint. The service runner measures recall, p50/p95/p99, errors and throughput at 1, 8 and 32 concurrent requests. An opt-in write test measures when updates and deletes become visible.
Local runs do not certify provider performance. A verified deployment report can be imported into the calculator’s decision panel alongside its cost estimate.
Open cost and decision toolEngineer’s handoff
- Download a workload or prepare an approved customer dataset.
- Load its record IDs, vectors and metadata into an isolated test namespace.
- Connect a search adapter and identify the exact provider, region and configuration.
- Run concurrent queries, review failures, then import eligible evidence.
Local commands & guide
npm run bench:local -- --all --profile standard npm run bench:service -- --fixture npm run bench:prepare-public
Full adapter contract and write-test safeguards: docs/service-benchmarks.md. Workload and measurement guide: docs/benchmark-lab.md.
Other strategies worth testing
HNSW + SQ / PQ / binary
Compressed graphs
Compression and graph routing can be combined. Test oversampling and full-vector reranking against your embeddings.
Primary reference ↗Sparse + dense + late interaction
Multi-stage retrieval
Use lexical and semantic retrieval to collect candidates, then rerank with a cross-encoder or multi-vector model. Measure relevance and end-to-end latency.
Primary reference ↗