How much does it cost to implement RAG: embeddings, storage, queries, and models

Published on 2026-11-16 by ArduBot

Cost model of a RAG service: initial indexing, OCR, embeddings, vector database, searches, generation, maintenance, and spikes.

A RAG prototype seems cheap when tested with three PDFs, but the cost changes if you process thousands of pages and receive queries all day. Before setting a monthly price, you need to separate indexing, retrieval, and generation.

The five costs worth budgeting

1. Extraction: downloading, converting HTML, or reading PDFs. Scanned pages may require OCR, which adds processing and error review.

2. Embeddings: transforming fragments into vectors. Billing may depend on the volume of text processed and the provider; reindexing updated documents generates new consumption.

3. Storage and indexing: storing text, metadata, vectors, copies, and versions in a database or a specialized engine.

4. Retrieval: search, filters, optional reranking, and server resources. Some solutions charge per query; others require paying for reserved capacity.

5. LLM response: if the service drafts responses, input and output tokens are an additional cost. An API that returns only fragments can operate without paying for generation on every search.

Numerical example: don't confuse assumptions with prices

Suppose 10,000 pages, 5 useful fragments per page and 20,000 monthly questions. We would have about 50,000 indexed units before duplicates or filters. If each query returns 5 fragments, the system will retrieve up to 100,000 fragments per month, but that does not mean 100,000 embedding calls: usually an embedding is calculated per question and embeddings per new content processed, according to the architecture.

To budget, multiply the estimated number of document tokens by the current rate of the embedding provider, add storage cost, CPU/ram, OCR, reranking and, if it exists, generation. Finally add a margin for spikes, retries and reindexation due to model changes.

There is no “universal RAG price” here. Provider rates can change and depend on the specific architecture; always request current quotes.

What influences more than it seems

A PDF with poorly extracted tables can produce useless fragments and force you to repeat the work. A product that allows uploading unlimited files can consume more in ingestion than in searches. Retention policies also matter: keeping versions indefinitely increases size and risk.

Cache can save repeated queries, but you must invalidate it when the source, permissions or authorized identity change. Never share a cache of private responses between clients.

A practical calculation template

Costo mensual =
  infraestructura fija
+ documentos nuevos/reindexados × costo de procesamiento
+ preguntas × costo de recuperación
+ preguntas respondidas × costo de generación
+ almacenamiento, backups, monitoreo y soporte
The commercial rate also needs to cover development, incidents, support, and margin, not just the provider's invoice.

How to measure the profitability of a RAG node

Offer interested parties a trial with a clear limit on documents and questions. Record setup time, monthly cost, and the client's actual operational savings. If manual support cost exceeds revenue, the service does not scale even if embeddings are cheap.

At ArduMaker we are evaluating an API-managed RAG, but we do not announce a price or availability before validating infrastructure and demand. If you are interested, indicate pages or GB, questions per month, and generation needs by visiting ArduMaker.

Sources for designing the architecture: RAG on Microsoft · AI API costs: related article.

  • RAG
  • costos IA
  • embeddings
  • SaaS
  • API

Keep reading

All articles · VPS plans