← Home

BrowserVec Examples

In-browser WebGPU vector store — learn every concept through runnable examples.

Most examples run on any browser via BrowserVec's CPU fallback; the quantization/IVF-focused examples (02, 03, 04, 09-worker-ingest, 10-chunking) require WebGPU.

Real-world use cases

• RAG RAG chat assistant Ingest documents, retrieve relevant context for LLM prompts — fully offline
• App Personal knowledge base Semantic note search with persistence — your notes survive reload
• Recs Movie recommender Find similar films by description — like "more like this"
• Dev Code snippet search Search code by intent, not keywords — "find the sort function"
• FAQ Customer support FAQ Semantic FAQ matching — users type a problem, get the right answer
• Ext Browser extension pattern Capture pages, store locally, semantic search — fully offline

Core concepts

• M1 Basic flat search fp32 brute-force GPU scan with cosine, dot, and L2 metrics
• M3 TurboQuant quantization int8 / int4 quantized storage with asymmetric query + exact re-rank
• M4 IVF approximate index cluster-based ANN, nprobe sweep, recall vs speed trade-off
• M4 IVF × quant combo clustered + quantized — the 1M-vector path with int8 or int4
• M4 IVF auto-tuning targetRecall picks nprobe for you — recall estimated on your own data at build time
• M7 HNSW graph index graph-based ANN, efSearch sweep, incremental inserts, works without WebGPU
• M7b GPU graph search CAGRA-style beam search in one dispatch — queryBatch, CPU vs GPU crossover
• React React hooks useVectorStore, useSimilaritySearch, useEmbedding, useRetriever — reusable React patterns
• Vue Vue composables useVectorStore, useSimilaritySearch, useEmbedding, useRetriever — reusable Vue patterns

Text & embeddings

• M5 Text retrieval end-to-end offline text→search with hashing embedder
• M5 Custom embedder implement the Embedder interface to plug in any embedding source
• M5 Transformers.js embedder real semantic embeddings via Xenova/all-MiniLM-L6-v2

Persistence & encryption

• M2 Persistence save/load to OPFS/IndexedDB, export/import as blob, cold-start reload
• M6 Encryption at rest AES-256-GCM encrypted snapshots with PBKDF2 passphrase

Data management

• M1 Delete & compaction tombstone deletes, update (upsert), compact(), over-fetch past tombstones
• M1 Metric comparison cosine vs dot vs L2 — same data, different scoring
• M1 Advanced API isSupported, stats, dimension config, destroy lifecycle

Performance & scale

• NFR-8 Worker ingest rotate+quantize offloaded to a Web Worker — no main-thread freeze
• NFR-10 Corpus chunking spread vectors across multiple GPU buffers past the per-buffer limit
• Bench Performance benchmark Large-scale perf test: ingest rate, query latency, recall across index types
• Bench Benchmark dashboard BrowserVec vs. hnswlib-wasm vs. voy-search vs. Orama vs. Vectorious vs. naive JS — latency, CPU/GPU/transfer breakdown, recall, memory (synthetic vectors)
• Bench Real-embeddings benchmark BrowserVec only (flat/int8/1-bit), real sentence embeddings (MiniLM, BGE, E5, Nomic) over 180 real documents, plus qualitative semantic search

Open each in a WebGPU-capable browser after running npm run dev.