Turn text into vectors and explore meaning as an interactive 3D map you can search by similarity.
Try this
- Search the demo corpus by meaning, not keywords, and watch the nearest matches light up.
- Open the Chunking tab to see how a strategy cuts a document into pieces before embedding — a separate demo from Search, not a live input to it.
- Rerank the results to sharpen relevance beyond raw vector distance.
Good to know
- Embeddings here are computed server-side by Google’s EmbeddingGemma, not in your browser.
- Points that sit close together mean similar things — distance is roughly semantic difference.
- MMR reranking is pure vector math — instant and deterministic, reordering the same dense candidates for relevance plus diversity. The LLM reranker (Retrieval tab) is a genuinely different, experimental mechanism: one real call to the self-hosted Gemma model, asked to judge relevance directly.
- RRF (Reciprocal Rank Fusion) fuses the dense and lexical rankings purely by rank position, not raw score — immune to one score’s scale dominating the blend, which is why it’s the safer 2026 default over min-max blending.
Ask Synapse
Still in training — may invent facts. Verify before you trust it.