Short answer: a focused RAG assistant over one well-organised source typically takes two to four engineer-weeks to build. A multi-source assistant with permissions, multiple channels and evaluation typically takes six to twelve engineer-weeks. Running costs are usually modest: embedding a document set is a one-off cost that is often small, and answering questions costs fractions of a cent to a few cents each depending on the model. These are typical ranges, not a quote.
Typical effort ranges
| Scope | Example | Typical build effort |
|---|---|---|
| Pilot | Q&A over a product manual or policy set, web widget | 2–4 engineer-weeks |
| Departmental | Several sources (wiki, drive, tickets), Slack or Teams bot, citations, evaluation | 5–8 engineer-weeks |
| Organisation-wide | Many sources, per-user permissions, SSO, multiple channels, analytics, self-hosted option | 8–14+ engineer-weeks |
Where the effort goes
- Content ingestion (often 30–40%). Connectors, parsing tricky formats (tables, scanned PDFs, slides), chunking strategies, metadata and incremental sync including deletions.
- Retrieval quality (20–30%). Hybrid search, re-ranking, filters, and testing whether the right passages are found.
- Answering and UX (15–20%). Grounded prompts, citations, refusal behaviour, conversation handling, channel integrations.
- Evaluation and launch (15–20%). Building a question set from real queries, scoring, pilot feedback and fixes.
- Security. Permission-aware retrieval and SSO add effort but are essential for internal knowledge.
Running-cost maths
One-off embedding. Suppose you have 5,000 documents averaging 2,000 tokens each: 10 million tokens. At a hypothetical embedding price of $0.02 per million tokens, that’s about $0.20. Even at ten times that price, embedding is rarely a significant cost. Re-embedding only changed documents keeps ongoing costs low.
Per question. A typical RAG answer sends a few retrieved passages plus the question to the model — say 4,000 input tokens and 300 output tokens. At a hypothetical $1 / $4 per million tokens, that’s about $0.005 per answer; at frontier prices perhaps ten times more. A thousand questions a day would then cost roughly $5–$50 per day in model usage. Always check current pricing.
Infrastructure. Vector storage, hosting and monitoring usually add a modest monthly fixed cost. PostgreSQL with pgvector keeps it simple for most organisations; dedicated vector databases make sense at larger scale.
Self-hosted. If data can’t leave your environment, add GPU server costs. That changes the economics — fixed monthly cost instead of per-question cost — and pays off at higher volumes.
Hidden costs to plan for
- Content quality. Outdated, duplicated or contradictory documents produce bad answers. Clean-up is often needed and is a content task, not an engineering one.
- Ownership. Someone should review unanswered questions and feedback weekly. That’s where most quality improvements come from.
- Change. New sources, new teams and new channels are natural extensions; plan for them.
Getting a reliable estimate
Share a sample of your content and a list of real questions people ask. In a short discovery we can test retrieval on your actual documents and give a fixed build price with expected running costs.