RAG Development Cost: What a Private Knowledge Assistant Costs to Build and Run
Short answer: A private knowledge assistant built with RAG (retrieval-augmented generation) typically costs $5,000–$12,500 to build for one team and a few hundred documents. Multi-source projects with permissions and a dashboard run $10,500–$26,500, and regulated builds $16,500–$41,000. Running costs for a small assistant can stay in the low hundreds of dollars a month.
Those build ranges come from our RAG cost estimator. Answer eight questions and it gives a range and timeline for your own project in about two minutes.
What you are paying for in a RAG project
RAG means the AI looks up your own documents before it answers. It does not rely only on what the language model learned in training. The assistant searches your files, picks the most relevant passages and writes an answer that cites them.
A production RAG system has five parts. Each one carries build effort and, for some, a monthly bill.
- Ingestion. Connectors pull documents from places like SharePoint, Google Drive, Confluence, a help centre or a database. Files are cleaned, split into chunks and kept in sync.
- Embeddings and index. Each chunk is turned into a vector (a list of numbers that captures its meaning) and stored in a vector database or search service.
- Retrieval. When a user asks a question, the system finds the best chunks. Good systems mix keyword and vector search and rerank the results.
- Generation. A language model writes the answer from the retrieved chunks, with citations, and says "I don't know" when the sources do not cover the question.
- Interface, permissions and monitoring. A chat window, Slack or Teams app, or API. Rules so people only see answers from documents they may open. Logs to check quality.
Most of the cost sits in steps 1, 3 and 5. Messy data, many sources and strict access rules take far longer than the model call itself.
Build cost: three typical scopes
We ran three common scenarios through the estimator's own pricing rules. The base price for a knowledge assistant is $4,000–$10,000 over 3–8 weeks. Data volume, integrations, a dashboard, usage and data sensitivity then move it up.
| Scope | What it includes | Build price | Timeline |
|---|---|---|---|
| Team knowledge assistant | One source, up to a few hundred files, works inside a tool you already use, under 1,000 questions a month, standard business data | $5,000–$12,500 | 4–9 weeks |
| Company-wide assistant | 1–2 connected tools, thousands of documents or messy data, admin dashboard, 1,000–20,000 questions a month, personal data | $10,500–$26,500 | 9–14 weeks |
| Regulated or multi-language assistant | 3–5 connected systems, large data set, dashboard, heavy use, two languages, health or finance rules or private hosting | $16,500–$41,000 | 11–16 weeks |
Optional support after launch is 15–25% of the build price a year, with a minimum of $200 a month. AI usage and hosting fees are extra and depend on volume. Urgent delivery in under three weeks adds 25% to the price.
What other firms publish
Published prices vary a lot, because firms define "RAG" and "production" differently. These are the figures we found on pages we opened this week.
| Source | Scope | Published figure |
|---|---|---|
| RaftLabs | Simple document Q&A | $12,000–$30,000 (4–8 weeks) |
| RaftLabs | Production, multi-source, with access control | $30,000–$60,000 (8–14 weeks) |
| RaftLabs | Enterprise platform | $70,000–$120,000+ (14–22 weeks) |
| Curiosity (illustrative in-house build) | Four engineers for one year, then one engineer to maintain it | About €600,000 to build, about €150,000 a year to maintain |
The Curiosity figure is a deliberately large example of a company building its own enterprise RAG platform from scratch. The company says the numbers are illustrative. It shows why most firms buy or commission a focused assistant instead.
For a wider view of agency rates, see our benchmark of 44 published AI prices.
Running costs: the monthly bill
After launch you pay for four things: the vector database, embeddings, language model calls and hosting. For most business assistants the monthly bill is small next to the build cost.
Vector database
Pinecone, a managed vector database, has a free Starter plan with up to 2 GB of storage. Its Builder plan is $20 a month flat. The Standard plan has a $50 a month minimum. Above that it charges $0.33 per GB a month for storage, $16–$18 per million read units and $4–$4.50 per million write units. Many projects start on the free or $20 tier. Self-hosting is another route (see hosting below).
Embeddings
Turning documents into vectors is cheap. OpenAI's embeddings guide says text-embedding-3-small processes about 62,500 pages per dollar, and text-embedding-3-large about 9,615, assuming around 800 tokens a page. On those figures, embedding 10,000 pages costs well under a dollar with the small model. You pay again only for new or changed documents.
Language model calls
This is the cost that grows with use. OpenAI's API pricing page on 5 November 2026 listed these standard prices per million tokens:
| Model (OpenAI) | Input per 1M tokens | Output per 1M tokens | Approx. cost of 10,000 questions* |
|---|---|---|---|
| gpt-6-luna | $0.10 | $0.50 | About $6.50 |
| gpt-6.1-terra | $1.00 | $5.00 | About $65 |
| gpt-6.1-sol | $2.00 | $10.00 | About $130 |
| gpt-6-astra | $10.00 | $50.00 | About $650 |
*Our own arithmetic, assuming each question sends 4,000 tokens of retrieved text and instructions and gets a 500-token answer. Longer context or chat history raises the cost. Cached input is cheaper on every model listed.
The lesson: model choice matters more than any other running cost. Test the cheaper models on your own questions first. Many internal knowledge questions do not need the most expensive model.
Hosting and monitoring
RaftLabs puts total infrastructure at $300–$2,500 a month, with most deployments at about 10,000 queries a month sitting at $400–$800. Its itemised table lists vector database, embeddings, model calls, app hosting, document ingestion and observability. It also says self-hosting can cut these costs by 40–60%.
What drives the price up or down
- Data quality. Clean PDFs and web pages are quick. Scanned files, tables, spreadsheets and duplicated versions need more work before they are useful.
- Number of sources. Each connector (SharePoint, Drive, Confluence, CRM, database) needs sync, error handling and testing.
- Permissions. If staff may only see answers from files they can open, the system must carry each document's access rules into search. This is one of the biggest jumps in effort.
- Languages and channels. English plus Arabic or Urdu, or website plus Slack or Teams, adds testing and interface work.
- Compliance and hosting. Personal or regulated data, audit logs and private hosting add time. See our AI for healthcare page for that sector.
- Accuracy target. A pilot that is "mostly right" is cheap. An assistant that staff or customers rely on needs a test set, reranking and regular reviews.
How to make answers accurate (and test it)
Most wrong RAG answers start with poor retrieval: the right passage was never found. Better search is often worth more than a bigger model.
Anthropic's research on "contextual retrieval" shows the size of the gain. Adding short context to each chunk before embedding cut the top-20 retrieval failure rate by 35%. Combining that with keyword (BM25) search cut failures by 49%. Adding a reranker cut them by 67%, from 5.7% to 1.9% in their tests. These are Anthropic's own results on its test sets, but the methods are widely used.
Before you sign off a build, ask for a simple evaluation:
- Collect 50–100 real questions from the people who will use the assistant.
- Write down where the correct answer lives for each one.
- Measure how often the right passage is retrieved, and how often the final answer is correct and cited.
- Check that the assistant refuses questions your documents do not cover.
- Re-run the same test after every change to data, prompts or models.
RAG, fine-tuning or an off-the-shelf tool?
RAG is the usual choice when answers must come from documents that change, and when you need citations. Fine-tuning changes how a model writes or classifies, but it does not keep facts up to date. Our guide to RAG vs fine-tuning explains the trade-off.
If you mainly need a website chatbot that answers customer FAQs, a simpler build may do. See how much a custom AI chatbot costs. If the assistant must also take actions, such as raising tickets or updating a CRM, it becomes an AI agent. The AI agent cost guide applies.
How to budget for a RAG assistant
- Pick one use case. For example, HR policy questions, sales enablement or support macros. Narrow scope keeps the first build in the lower range.
- Audit the data. List the sources, file types, volume and who may see what.
- Run the estimator. Our estimator turns those answers into a build range and timeline.
- Price the running costs. Use the model table above with your expected number of questions a month.
- Agree an accuracy test. Make the evaluation set part of the contract, not an afterthought.
Our custom knowledge retrieval service covers this full process, and the generative AI services page shows related builds. When you are ready, book a call and we will review your data and scope.
Frequently asked questions
How much does RAG development cost?
Our estimator puts a knowledge assistant at $4,000–$10,000 for the base build. A typical team project with some documents comes to $5,000–$12,500. Multi-source projects with permissions run $10,500–$26,500. Other firms publish $12,000–$120,000+ depending on scope.
What are the monthly running costs of a RAG system?
For a small internal assistant, model calls can cost a few dollars to a few hundred dollars a month, depending on the model. Add the vector database (free to $50+ a month on Pinecone) and hosting. RaftLabs puts most deployments at 10,000 queries a month at $400–$800 in total.
How long does it take to build a RAG assistant?
A focused assistant on one source takes about 4–9 weeks. Company-wide assistants with several sources, permissions and a dashboard take about 9–14 weeks, and regulated projects longer.
Is RAG cheaper than fine-tuning?
Usually, for question answering over company documents. RAG keeps answers current when documents change and shows its sources. Fine-tuning suits tasks where the model's style or classification must change.
Can a RAG assistant respect document permissions?
Yes, if it is designed for it. Each document's access rules must be stored with its chunks and applied at search time. This adds effort, which is why permissions move a project into the higher price band.
Sources
- RixDigi: AI cost estimator (knowledge assistant base range and pricing rules, 5 November 2026)
- RaftLabs: How much does RAG development cost? (build ranges and monthly infrastructure costs)
- Curiosity: What building your own RAG stack actually costs (illustrative in-house figures, July 2026)
- Pinecone: pricing
- OpenAI: API pricing
- OpenAI: embeddings guide (pages per dollar)
- Anthropic: Introducing contextual retrieval
