Real-World RAG Embedding Cost Examples: 5 Production Scenarios (September 2026)

On the embedding side, very little. In the scenarios above, the internal knowledge base (10k pages, 500 queries/day) costs $0.08-$0.24 to index once and under $0.03/month in query embeddings. The semantic search engine (1M documents, 10k queries/day) costs $5-$30 to index once and $0.18-$0.54/month in query embeddings. The vector database is usually the bigger recurring line: managed plans carry minimums (Pinecone Standard $50/month, Weaviate Cloud Flex $45/month) that exceed the embedding bill at these sizes, while pgvector on an existing Postgres database adds almost nothing. LLM generation for answers is separate and usually larger than both.

LLM generation cost is almost always higher for production RAG applications. Embedding a query costs about $0.0000006 (0.6 millionths of a dollar) on OpenAI small at 30 tokens. Generating a 500-token answer with Claude 3.5 Sonnet costs about $0.0015. The LLM is roughly 2,500x more expensive per request. Embedding optimization matters most for large indexing workloads, not query-time costs.

For a small knowledge base: OpenAI text-embedding-3-small for embeddings (cheapest commercial API), pgvector on an existing Postgres database (near-zero storage cost), and a local or smaller LLM for generation. Total embedding cost for a 100k-document knowledge base with 100 daily queries is under $2/month.

Only when documents change or when you upgrade embedding models. Static document sets (legal contracts, product manuals) rarely need re-embedding. Frequently updated content (news, tickets, knowledge base articles) should be re-embedded on update. The re-embedding cost is usually small relative to the initial indexing pass.

For large corpora, usually the database. Embedding 1 billion tokens with OpenAI text-embedding-3-small costs $20 once. At 500 tokens per chunk that is 2M vectors, about 11.4 GB at 1,536 dimensions, so Pinecone serverless storage at $0.33/GB comes to about $3.78/month. That is more than the Builder plan's 10 GB, and on Standard the $50/month minimum is the real floor, so the database bill passes the one-time embedding cost in the first month. Storage alone would pass it in month 6. Loading the vectors also uses at least 12.3M write units ($49-$55 once at Standard rates, before IDs and metadata). Read units for queries are extra and depend on traffic, which this example does not assume. pgvector or self-hosted Qdrant keeps this cost lowest.