What a production RAG pipeline actually costs to run
Retrieval sounds cheap until you price embeddings, storage and re-indexing at real volume. Here is the arithmetic we use before quoting a client.
1 min read
Every RAG proposal we review underestimates one line item: re-indexing. The first index is a one-off cost. Keeping it current is a subscription you pay forever.
Where the money actually goes
Embedding a 50,000-document corpus is a rounding error. Re-embedding it every time the source changes is not. We model three things separately: initial ingest, steady-state drift, and query volume.
The result is usually that storage and queries are trivial, and the entire bill is churn. That reframes the architecture — you optimise for cheap partial updates, not cheap reads.
