Why Your Vector Database Bill Is Too High (and the AWS Fix)

If your RAG project is paying a monthly vector database fee before it has real traffic, you are funding idle capacity. I made two videos on this, one explaining what vector search actually does and one building a full RAG pipeline on AWS for pennies, and this post pulls both together with the cost math so you get the substance without watching.

The one-minute version of vector search

Traditional search with BM25, stemming, and fuzzy matching is powerful, but it has one blind spot: it matches words, not meaning. Someone searches "affordable flights to Tokyo" and your content says "cheap airfare to Japan," and lexical search misses it. Vector search catches it. An embedding model turns text into a list of numbers, 384 dimensions or 1,536 or more, and similar concepts land close together in that space while "flight simulator video game" lands far away. A query finds the nearest points, usually measured by cosine similarity. Three parts carry it: the embedding model, the vector database, and the similarity metric. That is the backbone of RAG, semantic search, recommendations, and code search.

The store is where the money leaks

The model and the query logic are rarely the expensive part. The vector store is. OpenSearch Serverless runs 175 to 350 dollars a month. Pinecone starts around 50 dollars a month. For a small dataset with little traffic, that is standing cost with almost nothing to show for it.

Pennies on S3 Vectors

In the second video I swap that store for Amazon S3 Vectors under Bedrock Knowledge Bases.

I dropped personal docs, a resume, and YouTube channel data into a regular S3 bucket, pointed a Bedrock Knowledge Base at it, and picked the Bedrock data automation parser because it handles the PDFs. On sync it chunks the docs, runs them through Titan Text Embeddings V2, and writes vectors into an S3 vector bucket that the Knowledge Base creates for me. Querying with Claude Haiku 4.5, "what was Saurav's last job" came back as software development engineer on the docs platform at AWS, with a citation I could click straight to the source. On cost, the Knowledge Base and default parser are free, Titan Embeddings V2 is 0.20 dollars per million tokens, and S3 Vectors bills for storage plus queries. The demo came in under a dollar. The one thing to watch is query cost at large scale, but for a small project starting out it stays in the couple-of-dollars range.