What drives the cost of a RAG project?
The cost of a RAG project depends less on the model than on the state of your documents, the number of systems to connect, access and data-protection requirements, and usage volume. The budget has two parts: one-off work such as discovery, data preparation, integration and evaluation, and running costs such as model usage, hosting, monitoring and content upkeep.

Topic
AI
Published
17 November 2026
Updated
17 November 2026
What does the cost of a RAG project consist of?
The cost of a RAG project is really two budgets: the work done once to build the system, and the costs that continue for as long as it runs. Mixing the two makes quotes hard to compare. A quote with a low build fee can hide an architecture that is expensive to run.
This article gives no figures. Model and platform prices change often, and a project budget follows its scope. What it does give you is the list of drivers and what makes each one larger or smaller, so you can ask the right questions when you request a quote.
What are the one-off costs?
Discovery and scope. Someone writes down who the assistant serves, which questions it should answer and from which documents. Every later item grows while the scope stays vague.
Data preparation. In most projects this takes the most effort. It covers collecting documents, removing old versions, turning scanned PDFs into text, making tables readable and splitting text into passages (chunking). The gap between a clean, current document set and a mixed archive built up over years is the largest gap in the budget.
Integration. The system has to connect to wherever the documents live: SharePoint, Google Drive, Confluence, a file server or an ERP. Companies operating in Türkiye often run local ERPs such as Logo, Mikro or Netsis, and whether data comes out through an API or a file export changes the size of the job. Each new source means another connector, another permission mapping and another update flow.
Access model. If everyone may see every document, the work is simple. If each user should only get answers from documents they are allowed to read, the permissions in the source system have to carry over into search. That is a separate piece of design and testing.
Evaluation. A test set is built from real questions and the system is measured against it. Each language needs its own questions. Turkish, for example, adds suffixes to words, which affects search, so Turkish content should be tested with Turkish questions. Our post on RAG evaluation explains the method. Removing this item from a quote looks like a saving, but it makes the cause of every live error more expensive to find.
Security and data-protection review. If the documents contain personal data, the review covers legal basis, retention and security measures under GDPR, and under KVKK, Türkiye’s personal data protection law (Law No. 6698), if you operate there. Article 12 of KVKK requires the controller to take the technical and organisational measures needed for an appropriate level of security. Article 9 sets conditions for transferring personal data abroad, which matters when a model provider processes data in another country. This section is not legal advice; do the review with your legal counsel.
What are the running costs?
Model usage. Language models are billed per token. The pricing pages of Anthropic, OpenAI and Google all price input and output tokens separately. In RAG, the retrieved passages travel to the model with every question, so the input is far longer than the question itself. Monthly spend is roughly three things multiplied: the number of questions, the text sent per question and the unit price of the model. For the current unit price, check the provider’s pricing page.
Embedding and indexing. Documents are converted to vectors so they can be searched. OpenAI’s pricing page lists embedding models as a separate item. The cost arises when documents are first processed and again at every update. For large archives that change often, re-indexing is a regular expense.
Vector database and hosting. The vectors sit in a vector database. pgvector is an open-source extension that adds vector similarity search to Postgres. For a team that already runs Postgres, it allows a start without a separate service. Choosing a dedicated vector database such as Qdrant, or a managed cloud service, adds a line item in exchange for scale and easier operations. The application itself also runs on a server or a cloud service.
Monitoring. Questions, retrieved passages, answers and user feedback are logged. Storing those logs and reviewing a regular sample takes people’s time.
Maintenance and content upkeep. This is the item most budgets forget. Documents change, new procedures appear and model versions are retired. Without a content owner, the assistant starts answering from stale information and trust falls quickly. Who maintains the system should be clear at the quote stage.
Which choices raise or lower the cost?
| Cost driver | Type | Makes it cheaper | Makes it more expensive |
|---|---|---|---|
| Discovery and scope | One-off | One team, one group of questions, written success criteria | Vague goal, many stakeholders |
| Data preparation | One-off | Current, text-based documents in one place | Scanned PDFs, content heavy in tables and drawings, conflicting versions |
| Integration | One-off | A single source system with an API | Many sources, legacy systems without an API, customised ERPs |
| Access model | One-off | Internal content open to everyone | Document-level access that differs per person |
| Evaluation | One-off | A ready list of real questions | Several languages, several document types, sensitive topics needing human review |
| Security and data protection | One-off | Content with no personal data | Special categories of data, cross-border transfer, sector regulation |
| Model usage | Running | Smaller model, short context, caching | Large model, long context on every question, long answers |
| Vector database and hosting | Running | pgvector on an existing Postgres | Separate clusters, multi-region setup, on-premises servers |
| Monitoring and maintenance | Running | A named content owner, automated update flow | Manual updates, fast-changing content, an archive nobody owns |
How do you cut cost without cutting quality?
Retrieval quality before model size. If the wrong passage is retrieved, a larger model will not fix it. Better chunking and reranking of results often let a smaller model reach the same quality. Reranking has a processing cost of its own, so measure the gain on your test set.
A smaller model where it is enough. On the pricing pages, smaller models have lower unit prices than larger ones. For simple factual questions a small language model may be enough. Some teams route easy questions to a small model and hard ones to a large model. Choose with your own test set, as described in our post on choosing an LLM for business.
Keep the context short. A wide context window is not a reason to send dozens of passages with every question. Fewer, better-matched passages cost less and make the answer sharper.
Use caching. According to Anthropic’s documentation, prompt caching reuses repeated prefixes of a prompt and reduces processing time and cost. Writing to the cache costs more than normal input, and reading from it costs considerably less. It suits assistants with long, fixed system instructions. The OpenAI and Google pricing pages also price cached input separately.
Run non-urgent work in batches. For jobs that need no instant answer, such as summarising or tagging documents, providers offer discounted batch processing. Check the conditions on the provider’s pricing page.
Start with one document set. In the first release, take the one set that attracts the most questions, not the whole archive. Data preparation shrinks, measurement gets clearer, and you gain real cost data for the second set.
Leave fine-tuning for last. When the knowledge sits in documents, training a model is usually an unnecessary line item. Our page on RAG vs fine-tuning explains the difference.
What should you prepare before asking for a quote?
With the following in hand, quotes arrive faster and can be compared with each other:
- Document count and formats. Roughly how many documents, in which formats (Word, PDF, scanned PDF, Excel, web pages), and how often do they change?
- Systems to integrate. Where do the documents live, do those systems have an API, and where will people use the assistant (Teams, web, intranet)?
- Languages. Which languages are the content and the questions in? Will a question in one language need an answer from a document in another?
- Expected volume. How many users, roughly how many questions a day, and are there peak hours?
- Access. May everyone see every document, or does access differ per person?
- Personal data. Do the documents contain customer or employee data?
- Who approves answers. Who owns the content, who corrects a wrong answer, and which topics require human sign-off?
- Sample questions. Twenty or thirty questions that users really ask describe the scope better than any brief.
Where should you start?
- Pick the one topic and document set that attracts the most questions.
- Count the documents in that set and note their formats and how current they are.
- Collect a short list of real user questions.
- Clarify personal data and access with your legal and IT teams.
- Ask for a quote that separates one-off and running items, and get the assumed usage volume for the model in writing.
At Corrplus we follow the same working model on every project. Discovery starts with a free first call, followed by our two-week AI discovery sprint. In design we settle scope, architecture, and the data and data-protection assessment, and prepare a written quote. Build runs in short cycles, each ending with a working release. In the run phase, monitoring and improvement continue, as a managed service if you prefer.
For details, see our RAG enterprise knowledge assistant, LLM integration and AI consulting services. Our pricing page explains how we price our work.
Want help with this?
Book a call to apply what “What drives the cost of a RAG project?” describes to your company. We will listen and agree the first step together.
Sources
- Anthropic, Pricing (Claude API documentation)
- Anthropic, Prompt caching
- OpenAI, API pricing
- Google AI Studio, Gemini API pricing
- pgvector, Open-source vector similarity search for Postgres (GitHub)
- KVKK (Turkish Personal Data Protection Authority), Personal Data Protection Law No. 6698
Written by the Corrplus Team. How we write and check our posts: editorial policy.
Frequently asked questions
Is there a typical price range for a RAG project?
No reliable single range exists, because cost follows the state of the documents, the number of integrations, the access model and usage volume. Two projects built on the same technology can end up with very different budgets. The sound approach is to write the scope down and ask for a quote broken into line items.
What does the monthly running cost of RAG depend on?
It depends on the number of questions, how much text is sent to the model with each question, the model you choose and the infrastructure. Model providers price input and output tokens separately, and prices change often. Check the provider's pricing page for current rates and measure with your own sample questions.
Does a larger model give better answers?
Not always. If the right passage never reaches the model, the largest model still cannot write the right answer. Measure and fix retrieval first, and move to a larger model only when retrieval works well and the answers still fall short.
How does a data-protection review affect the cost?
If the documents contain personal data, the legal basis, access model, log retention and the country where data is processed need to be settled at the start. That work takes time in the design phase and can change the architecture. Do the review with your legal counsel.
Related services
Related articles
How to keep an AI agent inventoryA method for tracking owner, data, permissions, model, cost and logging for every AI agent as their number grows.
MCP 2026-07-28: a stateless coreWhat MCP 2026-07-28 changes: no sessions, routing headers, deprecated features and a migration checklist for teams running MCP agents.
AI shopping assistants and Black FridayBlack Friday prep for e-commerce teams: the product data AI assistants read, and customer service that holds up at peak.Let’s talk about your transformation
Tell us which work repeats every week. We will tell you honestly whether it is worth automating.
How we start
- A free first callWe pin down the problem and the goal together.
- A 2-week discovery sprintWe look at the data, the process and the tools, and decide where to start.
- A written proposalScope, architecture and timeline, in writing.