How much does a production RAG assistant cost?

What you are actually paying for

A retrieval-augmented generation (RAG) assistant is not "a chatbot with an API key." A production system has four parts, and each one costs engineering time:

• Ingestion — crawling or importing your documentation, chunking it sensibly, and embedding it into a vector store. This has to be re-runnable as your docs change.
• Retrieval — turning a user question into the right chunks of your content, fast, with relevance that holds up on real queries rather than demos.
• Generation — composing a grounded, cited answer with schema-constrained output so the model cannot return malformed or unsupported responses.
• Operations — monitoring, prompt and index tuning, model updates, and reporting. This is where most "AI projects" quietly fail after launch.

The real cost drivers

Two systems answering the same kind of question can differ 5x in price. The drivers are rarely the model:

• Documentation volume and messiness — clean Markdown is cheap; a decade of PDFs, tickets, and wiki sprawl is not.
• Accuracy and safety requirements — a marketing FAQ tolerates a wrong answer; a clinical or financial assistant needs deterministic guardrails around the model.
• Integration depth — a standalone widget is far cheaper than an assistant wired into your auth, your product data, and your support tooling.
• Ongoing token and embedding cost — usually small (embeddings can run near-zero on a free tier), but it scales with traffic.

Build vs buy

Off-the-shelf widgets are cheap and fine for a simple public FAQ. They stop working the moment you need your own auth, your own data, grounded citations, or guarantees about what the model can and cannot say. A custom build costs more up front and pays back in control: you own the pipeline, the data model, and the accuracy.

The honest rule of thumb: if a wrong answer is a minor UX annoyance, buy a widget. If a wrong answer costs you a customer, a compliance issue, or a support escalation, build it properly.

A realistic budget

At Khaas the ladder is deliberately simple: a €2,500 one-week Discovery Sprint scopes the highest-ROI use case and estimates ROI per case; a production Support Deflection System is a fixed-scope build from €18,000; and an AI Care Retainer keeps it accurate from €1,500/month. The sprint fee is credited to the build if you proceed, so you commit with the numbers in front of you rather than a vendor guess.