Initializing LLM...
First-time model download may take a moment
Real WASM Demo

WarpVector Playground

See how vector space transforms in real-time using the actual warpvector library with WASM acceleration. Every dot movement is a real IntentAdapter.tune() call.

πŸ‡―πŸ‡΅ ζ—₯本θͺžη‰ˆ

⬑ Vector Space 32-dim β†’ 2D
Loading… Vanilla Search
Document Query Top Match
β€”
Transform Latency
β€”
Top-1 Similarity
β€”
Rank Improvement
Live API Call
// No intent applied
const result = baseVector;
⚑ Optimization & Compression
1. Whitening (De-bias)

Fixes embedding anisotropy. Spreads vectors for higher resolution.

2. Quantization 1536 Bytes/vec

Compresses vectors. Watch rankings remain stable despite massive size reduction.

ROI Simulator (1M vectors) $0 saved
Estimated DB Storage: 6.0 GB
Est. Monthly Cloud Cost: ~$180 / mo
β—ˆ Search Intent

Add dynamic intent: Generates an LLM vector from string and creates a new transformation matrix.

≑ Search Rankings
⚑ Batch Performance
1,000 vectors Γ— 32-dim
πŸ€– Auto-learn Intents
IntentMatrixFactory

Uses IntentMatrixFactory to automatically learn optimal intent matrices from the document categories using InfoNCE contrastive learning.

✦ How It Works

This demo runs the actual warpvector library in your browser:

  1. 15 documents are generated as Float32Array[32] vectors
  2. IntentAdapter is initialized with real affine matrices
  3. Clicking an intent calls .tune() or .tuneBlended()
  4. WASM tuneBatchWasm() accelerates matrix multiplication
  5. Transformed vectors are projected to 2D via Gram-Schmidt basis