Vectors Wizard

Explore 13 vector database cost models, build an evidence-based shortlist, and inspect real HNSW/IVF search traces. Built by Caylent.

Live demo: vectors-wizard.caylent.tech

September 2026 refresh (local working tree): current billing rules and sources, native DynamoDB vectors, OpenSearch NextGen, S3 partitioning, Qdrant/pgvector capacity quotes, infrastructure and operating budgets, independent cost curves, benchmark import and decision-brief export, and corrected algorithm visualizations. See AWS capabilities, infrastructure costs, pricing evidence, and visualization audit. The live demo has not been updated by this local review.

Supported Providers

Provider	Pricing Model
Amazon S3 Vectors	Partitioned storage + batched PUT + query fan-out + processed/returned data
DynamoDB native vectors	Base table + vector index; measured search bytes required for ranking
OpenSearch Classic	OCU-hours + managed storage
OpenSearch NextGen	Active OCU-hours + hot storage + optional GPU acceleration
Pinecone	Serverless read/write units, storage, and plan minimums
Weaviate Cloud	Per-dimension pricing + SLA tiers
Zilliz Cloud	CU-based compute + storage
TurboPuffer	Storage + read/write GB + plan minimums
MongoDB Atlas	Cluster + Search Nodes + sharding + operating budget
MongoDB on EC2	Database + mongot + EBS performance + operating budget
Milvus on EC2	EC2 + EBS + dependencies/object storage + operating budget
Qdrant	User-supplied sized capacity quote + operating budget
PostgreSQL / pgvector	User-supplied deployment estimate + operating budget

Features

• Guided Wizard — chatbot-style questionnaire that walks through use case, embedding model, and workload parameters

  

• Manual Configurator — direct control over all pricing inputs with real-time cost breakdown

  

• Cross-Provider Comparison — 13 models with preserved workloads; unavailable or unpriced configurations are excluded from ranking

  

• Quickstart Presets — pre-configured templates for common use cases (RAG chatbot, product search, knowledge base, image similarity)

• Shareable Links — URL-encoded state for sharing configurations

• Import/Export — save and load configurations as JSON

• 3D Visualizer — interactive Three.js visualizations of HNSW and IVF index algorithms

  

• Semantic Map — a 2D map at /visualizer/map with local text embeddings, worker-based projection/index/search, larger synthetic datasets, and exact vs hierarchical HNSW vs trained IVF comparison

• Decision Workbench — budget/recall/latency/traffic targets, validated benchmark JSON import, matching against workload and capacity, and a Markdown decision brief

• Cost Curves — vary stored vectors or monthly searches independently, inspect sampled costs and bracket crossover ranges

• Capability Guide — /capabilities explains deployment options, filtering, hybrid retrieval, quantization and the evidence to collect

Interpreting the Results

Prices are estimates using the modeled region, plan, usage, and allocation assumptions. Cross-provider translation preserves a common workload, but does not establish equivalent capacity, latency, availability, or recall. Check provider-specific research under research/ and the provider's official pricing before making a purchasing decision.

The algorithm lab executes deterministic hierarchical HNSW insertion/search and trained IVF-Flat, showing actual traces and recall against exhaustive ground truth. The 3D scene uses small synthetic 3D fixtures; hierarchy shells are a schematic layout, not vector distances. Map timing includes trace capture on the current device and is not a vendor latency benchmark. Byte/RAM sizing remains an explicit estimate. Text projections and first-two-coordinate views can distort higher-dimensional distance. Imported benchmark reports are user-supplied evidence, not independently verified measurements; a passing assessment covers only the entered targets, not production readiness.

Quick Start

npm install
npm run dev

Open http://localhost:3000

Architecture

Built with Next.js 16, React 19, TypeScript, and Tailwind CSS. The calculator uses React hooks; the visualization and embedding features use Zustand stores.

Provider System

The core is a pluggable provider abstraction in src/lib/providers/. Each provider implements PricingProvider<TConfig> and defines:

• Pricing logic — rates and cost calculation (pricing.ts)
• Config fields — UI form definition (index.ts)
• Wizard steps — guided questionnaire flow (wizard-steps.ts)
• Presets — example configurations (presets.ts)
• Cross-provider translation — toUniversalConfig / fromUniversalConfig for comparison

State Management

Two custom hooks manage calculator state:

• useCalculator — config, mode switching (landing/wizard/configurator), cost computation via useMemo
• useWizard — conversation flow, step tracking, branching logic, config patching

Zustand stores in src/stores/ manage index parameters, visualization playback, embedding sessions, and map controls. useSemanticMap coordinates local embedding generation, projection, in-session caching, and neighbor comparisons.

Visualizer

Interactive 3D visualizations at /visualizer using @react-three/fiber:

• HNSW — hierarchical navigable small world graph with animated search traversal
• IVF — inverted file index with centroid-based nearest neighbor search

Additional routes:

• /visualizer/embed — generate text embeddings in a browser worker
• /visualizer/map — inspect local neighborhoods on an SVG map with text projections from umap-js
• /visualizer/compare — compare index structures and qualitative tradeoffs
• /capabilities — service capabilities, cost-model coverage and evaluation questions

The map includes 48 original support prompts, 24 adapted news-style examples, 800 seeded geometry vectors, and 2,000 generated 32-dimensional Fourier-ring vectors. Synthetic generators and provenance are documented in the UI. Text models download on first use. Projection, index construction and search run in a cancellable worker; indexes are cached across query changes. These are learning datasets, not representative production benchmarks.

Development History

The project began as an S3 Vectors calculator on February 4, 2026. Multi-provider comparison and the 3D visualizer followed on February 5, then pricing, input-validation, accessibility, and animation fixes on February 18. A March 6–9 Codex session added the Apple Embedding Atlas-inspired semantic map to the local working tree.

See docs/build-history-review.md for the evidence, original design choices, remaining gaps, and proposed next steps. The earlier docs/visualizer-improvement-plan.md records the 2D analysis and search-playback roadmap.

Scripts

npm run dev      # Development server
npm run build    # Production build
npm run lint     # ESLint
npm test         # Tests in watch mode
npm run test:run # Run tests once

License

MIT

Workload Lab

Open /benchmarks for eight reproducible stress workloads plus a real SQuAD/MiniLM passage retrieval fixture. Compare exact scan, HNSW, IVF-Flat, IVF-SQ8, IVF-PQ, LSH and dense/BM25 fusion with held-out ground truth, relevance labels, local timings and estimated index payload. Export datasets and reports, or connect the service harness for concurrency and opt-in freshness checks.

npm run bench:local -- --all --profile standard
npm run bench:service -- --fixture

See Workload Lab guide, service benchmarking, dataset provenance, and index strategy research. Local algorithms and protocol fixtures are explicitly separate from provider benchmark evidence.