Two vendors can quote $20K and $200K for the same sentence: "we want a chatbot over our documents." Both can be honest, because they priced different projects. These eight decisions tell you which project you have.
RAG stands for retrieval-augmented generation. The system first finds the right passages in your documents, then a language model answers from them. The idea fits in one sentence. The project does not.
This post is for the person who has to write the brief. Settle the decisions below before you ask for quotes, and vendors will price the same project. You will also spot the quote that left something out.
Figure 1 puts published price ranges on one scale. They come from vendor blogs, not independent data, and each vendor draws its tiers differently. The spread is the point: the bottom of the first bar and the top of the last are 50 times apart.
The language model is rarely the reason. Tensoria's breakdown puts model API calls at 5% to 15% of the budget, and engineering and data preparation at 50% to 60%. It also describes how "can we also add..." requests turn a $45,000 project into a $90,000 one. Most of the money follows scope.
If your knowledge base is small, you may not need retrieval. Anthropic's guidance says that under about 200,000 tokens, roughly 500 pages, you can put everything in the prompt and let prompt caching keep it fast and affordable. That advice dates from 2024, and limits differ by model, so check it against the model you plan to use.
If the system must act, not only answer, you are scoping something else. Our post on RAG pipelines vs AI agents covers that choice.
Figure 2 shows both ends of each decision. Most real projects land somewhere between them.
Name one user group for the first release. Then collect 30 to 50 questions they really asked, from tickets, emails or chat logs. If nobody can list them, that is your first task. These questions decide everything below, and they become your test set.
Pick the fewest sources that answer those questions. Each extra source type needs its own connector. Tensoria prices a connector at $5,500 to $16,500, and says teams that plan two days per system integration often spend two to three weeks. Write down what stays out of the first release.
Open a sample of 50 documents and look. Are there scans, complex tables, duplicates, or three versions of one policy? Tensoria says messy collections can multiply engineering time by 3 to 5. DevStudio calls ingestion and cleaning the most underestimated cost. Also name an owner for each source. Someone has to decide which version is true.
If HR files and sales notes share one index, can a new hire pull up salary data by asking? OWASP lists this under LLM08:2025 Vector and Embedding Weaknesses and recommends permission-aware stores with strict partitioning between groups. Carrying each source's access rules into retrieval is real engineering. In our view it is much cheaper to design in than to add later.
A nightly reload is a simple job. Showing changes within minutes needs event-driven sync. Deletions matter as much as updates, because a withdrawn policy that still answers questions is a real failure. Set one target per source.
Must every answer cite its source? What happens when the evidence is thin: say "not found", or hand off to a person? Which languages? Each choice adds design and test work, so decide now and not in week six.
Turn your real questions into a test set. Braintrust suggests starting with 30 to 50 real or realistic queries, including ambiguous and unanswerable ones. Score retrieval and the final answer separately, so a bad answer can be traced to its cause. Agree the pass marks before the build starts.
This is also how you judge technique. In Anthropic's tests, contextual retrieval plus reranking cut top-20 retrieval failures from 5.7% to 1.9%. Those were their datasets, and yours will differ. Without a test set, you will never know by how much.
Build cost is only part of the bill. Tensoria estimates first-year total cost at 1.5 to 2 times the build, with maintenance at 15% to 20% of the build each year. DevStudio lists retrieval tuning and support at $1,000 to $5,000 or more a month. Decide who updates content, who reviews failed questions, and who owns the budget.
Your brief does not need to be long. One page should state:
Send the same page to every vendor. Ask each to list assumptions and exclusions next to the price. For more questions to ask, see our checklist for choosing an AI development company.
Do not sign for the whole platform on day one. Start with one user group, one or two sources and your test set. Check the scores against your pass marks on real documents. Then add sources and users. A demo on ten clean PDFs proves little, which is why so many proofs of concept never ship.
Our AI project cost estimator starts a knowledge assistant at $10K to $25K over 4 to 10 weeks. Integrations, messy data, compliance and private hosting raise that, and the page shows every assumption. These are planning figures, not averages from past work. Innometrique was founded in 2025 and does not have a long delivery history to average.
Not sure your data is ready? The AI readiness assessment takes a few minutes. If you cannot settle these decisions in-house, a short paid discovery can do it on your real documents. It is listed on our pricing page.
Send us your one-page scope, or just the questions your team keeps asking. We will tell you plainly what is missing and whether you need RAG at all.
Talk to Us →