Retrieval-augmented generation (RAG) lets an AI model answer questions using your own documents, policies and data instead of only what it learned in training. It's the most common way businesses build accurate AI assistants today. But budgets vary widely, so here's how to think about cost.
What goes into a RAG system?
Every RAG system has the same core parts: a pipeline that ingests and cleans your content, an embedding step that turns it into searchable vectors, a vector database, a retrieval layer that finds the right passages, and a language model that writes the final answer. Around that sit the user interface, access control, monitoring and evaluation.
Typical cost ranges
| Scope | Typical range | Timeline |
|---|---|---|
| Proof of concept on a small document set | $8,000 – $20,000 | 3–5 weeks |
| Production assistant for one team | $25,000 – $70,000 | 6–10 weeks |
| Enterprise platform with many sources and roles | $80,000 – $200,000+ | 3–6 months |
These are planning ranges, not quotes. The final number depends on the factors below.
The five biggest cost drivers
- Data preparation. Messy PDFs, scanned files and spreadsheets take far more work than clean text.
- Number of sources. Each connector (SharePoint, Google Drive, a CRM) adds build and testing time.
- Accuracy requirements. Legal, medical or financial answers need stronger evaluation and human review.
- Security and permissions. Respecting who can see which document adds complexity, but it's essential.
- Usage volume. Model and hosting costs grow with the number of questions asked each month.
Ongoing running costs
After launch, expect monthly costs for model usage, vector database hosting and monitoring. For most mid-sized deployments these are modest compared with build cost, but they scale with traffic, so it's worth caching common answers and choosing the right model for each task.
How to keep your budget under control
- Define success metrics before you build, such as answer accuracy or time saved.
- Clean and structure your most important content first.
- Use smaller, cheaper models where they perform well enough.
- Build evaluation in from day one so you catch problems early.
If you're exploring a RAG project, we're happy to review your use case and give you a clear estimate.
