VectorLens

See what your embeddings actually look like — project, cluster and search them.

Embeddings encode your source documents — these never leave this tab

Embeddings — JSON, JSONL or CSV

PCA projection

Click a point to pin it — its five nearest neighbours are linked and everything else dims.

Similarity search

Accepted formats

Paste or open any of these — the format is detected automatically:
  • [{"id": "…", "text": "…", "embedding": [ … ]}] — the usual export shape. vector, values and vec work too, as do label/name/title for the caption and content/chunk/page_content for the text.
  • {"data": [{"embedding": [ … ]}]} — an OpenAI embeddings response, pasted straight in.
  • [[0.1, 0.2, …], [ … ]] — bare arrays of numbers.
  • JSONL — one JSON object per line, as most vector-store exports write it.
  • CSV/TSV — numeric columns form the vector; the first text column becomes the label.
Where this fits with a real vector database. Pinecone, Weaviate, Qdrant, Milvus, Chroma and pgvector all export embeddings as JSON or CSV — that export is what you drop in here. The tool works on the vectors themselves, so it is not tied to any one store, and no connection or API key is involved.
Read the projection carefully. PCA squashes hundreds of dimensions onto two, so it always loses information — the "variance shown" stat tells you how much survived. Cosine scores in the table are computed on the full vectors and are exact; the picture is the approximation, not the numbers.