How Much Does a RAG Chatbot Cost in 2026? Build Price, Running Cost, and the Cheaper Route

A RAG chatbot, one that answers questions from your own documents, costs $10,000–$25,000 for a proof of concept and $30,000–$80,000 for a production system at typical 2026 agency rates. Monthly running costs at small-business volume are $200–$1,500. Most of the build cost is data preparation and evaluation rather than the model, which is why a single experienced engineer can ship a focused first version in three to four weeks for far less.

Retrieval-augmented generation (RAG) is the technique behind almost every "ask our documents" system: when a question comes in, the system finds the most relevant passages from your files, hands them to a language model, and the model writes an answer grounded in those passages. It is the most common first AI project for a business, and the most commonly over-quoted. This guide breaks down where the money actually goes.

What a RAG system costs to build

The gap between the agency and Wope rows is not magic. It is three things: one senior engineer instead of a team with a project manager, a scope that starts narrow, and an hourly rate set by where the engineer lives rather than where the client does. See what drives fractional AI engineer pricing for the full breakdown.

Where the build money actually goes

1. Data preparation (30–40%). Collecting the documents, converting scanned PDFs, deduplicating, splitting into chunks that keep their meaning, and resolving the fact that your 2019 refund policy contradicts your 2024 one. This is the part that decides whether the system is trustworthy and the part most quotes leave out.
2. Retrieval design (15–20%). Choosing how documents are indexed and searched: embeddings, keyword search, or both; how many passages to fetch; how to rerank them. Getting this wrong is why some bots confidently answer from the wrong document.
3. Evaluation (15–20%). A test set of 100–200 real questions with known-good answers, run on every change. Without it you cannot tell whether a tweak helped. With it, accuracy regressions fail the build before anyone sees them.
4. Integration and interface (15–20%). Wiring it into the help desk, Slack, the website widget, or wherever the questions actually arrive, plus the human-handoff path.
5. Model and prompt work (10%). Smaller than most people expect. The model is the commodity part.

What a RAG chatbot costs to run per month

Some reference points. Embedding documents is close to free: with a current small embedding model at roughly $0.02 per million tokens, indexing 10,000 typical documents costs on the order of ten cents. A managed vector database for a small or mid-sized corpus runs from tens of dollars to a few hundred a month. The language model that writes the answer is where the monthly bill lives, and it scales with query volume and answer length. A support bot handling 3,000 questions a month will usually land under $300 a month in model costs.

RAG vs fine-tuning: which should a business choose?

RAG, for almost every case where the goal is answering from company knowledge. The reasons are practical:

• Freshness. When a policy changes, RAG re-indexes one document. A fine-tuned model has to be retrained.
• Traceability. RAG can show which passage the answer came from. A fine-tuned model cannot, which matters the first time it is wrong about a contract clause.
• Cost. Fine-tuning runs are a recurring cost every time the knowledge changes; RAG indexing is pennies.
• Access control. RAG can filter which documents a given user is allowed to retrieve. Fine-tuning bakes everything into one model for everyone.

Fine-tuning earns its place when you want to change how the model writes rather than what it knows: a very specific output format, a house tone across thousands of drafts, or a narrow classification task at high volume. Many production systems eventually use both. Start with RAG.

What a real small-business RAG project looks like

A worked example, drawn from the kind of engagement we run most often. A 40-person services firm, support team of four, roughly 900 tickets a month, of which about 60% are repeats of a few dozen questions.

• Week 1: three years of tickets and the help-centre articles ingested and deduplicated. The engineer finds that 11% of "resolved" tickets were closed by timeout and excludes them from the gold set. A test set of 120 real questions with answers the support lead approves.
• Week 2: retrieval built and tuned against the test set. First commit in the client's repository. Accuracy scoring wired into their pipeline so any change that lowers the score fails the build.
• Week 3: live behind a flag to 10% of incoming tickets, with a confidence threshold below which it hands off to a human. 41 of the first 52 answered without help.
• Week 4: widened to 50% of tickets. Support time on repeats down from 62 hours a month to under 25.

Build cost: one month on the Core plan, $2,100. Running cost: roughly $180 a month in model and database fees. The saving, at a loaded support cost of $35 an hour, is around $1,300 a month, so the system paid for itself in its second month.

Questions to ask before you accept a RAG quote

1. How will you handle contradictions between old and new documents?
2. What is in the test set, who approves it, and when does it run?
3. Where does the answer go when confidence is low?
4. Which documents can which users retrieve?
5. What is the monthly running cost at our volume, in dollars?
6. Which repository and cloud account will this live in?

A firm that answers all six without hesitation has built this before. A firm that cannot is quoting you for its own education.

Frequently asked questions

How much does it cost to build a RAG chatbot?

Published 2026 agency prices are $10,000–$25,000 for a proof of concept, $30,000–$80,000 for a production system and $80,000–$200,000+ for enterprise scale. A focused first version built by one senior engineer over three to four weeks costs a fraction of that: at Wope, it is the first month's work on a Core plan.

What does a RAG chatbot cost to run per month?

For up to 10,000 queries a month, roughly $200–$1,500 in model, embedding, vector database and hosting costs. Embeddings are nearly free; most of the bill is the language model answering questions.

RAG or fine-tuning: which is cheaper for a business?

RAG, almost always. Fine-tuning bakes knowledge into a model and has to be redone when documents change; RAG looks documents up at answer time, so updating is just re-indexing. Fine-tuning is for changing a model's style or format, not for teaching it your policies.

How long does it take to build a RAG system?

A production-grade first slice over one document set takes three to six weeks with an experienced engineer. Most of that time is data cleanup and evaluation, not the model.

What is the hidden cost in RAG projects?

Data preparation. Deduplicating, chunking and resolving contradictions between old and new documents can be 30–40% of the budget, and it is the part most quotes leave out.