The short answer
A RAG knowledge assistant costs $12,000 to $45,000 to build with Lumeor and takes three to eight weeks. The low end is one clean document library that everyone can read. The high end is several sources, per-user permissions, messy file formats and compliance requirements.
The question that moves the price most is permissions. If everyone in the company may see every document, the build is much simpler. If HR files, contracts or client folders are restricted, the assistant must enforce the same rules your file system does.
If your content is a public help center and the audience is customers, you probably want a lighter AI chatbot instead.
Where the money goes
| Phase | Share of effort | What happens |
|---|---|---|
| Source audit | 10–15% | Find where the knowledge really lives, decide what is in and out of scope, check access |
| Ingestion and parsing | 20–30% | Connectors, OCR, table handling, chunking strategy, metadata |
| Retrieval tuning | 15–20% | Embeddings, hybrid keyword and vector search, reranking, permission filters |
| Answer generation | 10–15% | Prompts, citation format, refusal when sources are thin |
| Evaluation | 15–20% | Graded question set, retrieval accuracy checks, fixes and retests |
| Interface and deployment | 10–15% | Slack, Teams or web app, sign-in, feedback buttons, monitoring |
Ingestion is the surprise for most buyers. A 40-page policy PDF with tables, headers on every page and scanned appendices can take more care than a thousand clean web pages.
What it costs to run
A RAG assistant has more moving parts than a simple chatbot, and each running cost scales with something different:
- Embeddings. Billed per token to convert documents into vectors. Large up front for big libraries, then small as only changed files are re-embedded.
- Vector database. A managed service such as Pinecone is billed on stored vectors and queries. pgvector inside a Postgres database you already run adds little beyond storage. See Pinecone vs pgvector for the trade-offs.
- Model usage per question. Billed per token. It scales with the number of questions and with how many retrieved passages are sent with each one, so retrieval tuning directly affects the bill.
- Reranking and OCR. Rerankers are usually billed per query. OCR services bill per page processed, mostly during the initial load.
- Sync jobs and hosting. Scheduled jobs that pull changed files, plus the app server. Small for most companies.
- Monitoring. Logs of questions, retrieved sources and feedback, so you can see where answers go wrong.
How to keep the cost down
- Index less. Archive folders and duplicate drafts make answers worse and cost money to embed. Start with the documents people actually ask about.
- Settle permissions early. Decide on day one whether the assistant needs per-user filtering. Adding it later means rebuilding the index.
- Reuse your database. If you already run Postgres, pgvector avoids another vendor at small and medium scale.
- Use Slack or Teams first. Meeting people where they already work avoids building a full web app.
When not to build one
If your knowledge base is small and lives in one suite with correct permissions, test the AI search built into that suite before paying for a custom build. If the real problem is that nobody has written the answers down, an assistant will not fix it. If you are weighing training a model on your files instead, read RAG vs fine-tuning first.
How Lumeor prices RAG projects
We start by sampling your documents and collecting 30 to 50 real questions people ask today. The fixed-price quote lists the sources, file types, permission model, interface and the question set we will grade against at sign-off. Everything runs in your accounts, so the index and the code stay with you. Our step-by-step guide to building a RAG knowledge assistant shows the architecture in more detail.