Semantic Search
Keyword search matches letters. Semantic search matches meaning. Type a question and watch real Amazon Titan embeddings rank documents by what they mean — even when the words don't overlap.
🔎 Meaning, not matching
Ask “how do neural networks pay attention?” and a good semantic search returns the sentence about self-attention in transformers — even though it shares almost no words with your question. That's because both were turned into vectors that sit close together in meaning-space. This is the retrieval half of RAG.
Free uses this fixed corpus. Sign in for 3 demo searches/day; Pro beta members can search their own documents.
How it works
From question to ranked results
Embed the query
Your question is turned into a 1024-number vector by Amazon Titan — a point in meaning-space.
Embed the documents
Every candidate document becomes a vector too, using the exact same model.
Compare by similarity
Cosine similarity (a dot product of normalized vectors) scores how close each document is to the query in meaning.
Rank & return
The closest vectors win — even if they share no keywords with your question. With OpenSearch, a k-NN index does this at scale.
Where the credentials live
All embedding calls happen on the server. AWS credentials are read from environment variables on the deployment and never reach your browser or the source code. The page only ever sees the ranked results.
Related reading · Knowledge Lab