Embedding Space Explorer

Questions

Embeddings and RAG, in plain terms

• What is a text embedding?

  A list of numbers that stands in for meaning. A model reads a piece of text and returns a fixed-length vector — 768 numbers here, from Google EmbeddingGemma running on this studio's own server — placed so that texts meaning similar things land close together and unrelated ones land far apart. That is the whole trick behind search by meaning: comparison stops being a language problem and becomes arithmetic.

• What is RAG, and what does this tool actually show of it?

  Retrieval-augmented generation: instead of hoping a model memorised a fact, you look the fact up and hand it to the model along with the question. This tool exposes every stage of the retrieval half — how text is cut into chunks, how those chunks are embedded, how a query is matched against them, and how the shortlist is reordered before anything is passed on. Those stages are where RAG systems actually fail in production, and in most products they are completely invisible.

• Why combine vector search with old-fashioned keyword search?

  Because they fail differently. Vector search understands that "car" and "automobile" are one idea, but it drifts on exact strings — a part number, a surname, an error code. BM25 keyword scoring nails those and is helpless with synonyms. Fusing the two rankings — Reciprocal Rank Fusion here — keeps each one's strength, which is why serious retrieval systems rarely rely on either alone.

• What is Matryoshka truncation, and what does it cost you?

  EmbeddingGemma is trained so that the front of the vector carries most of the meaning. That lets you cut 768 numbers down to 512, 256 or 128 and still search usefully — a quarter of the storage, and noticeably faster comparisons. What you pay is precision, along a curve you can watch here instead of reading about: the same search re-runs at each width, so you can see where quality actually starts to break for your data rather than trusting a benchmark run on someone else's.