Chat with your docs (RAG knowledge base): what to build it with and what it costs
Upload documents and ask questions about them; answers cite the passages they came from.
Sample PRD
- Document upload
- Search over passages
- Answers with citations
- Team workspaces
What it needs
- Text generation (LLM API): a language model writes answers from the passages
- Postgres: passages and embeddings (pgvector) in Postgres
- Organisations / B2B teams: each team has its own documents
- Subscriptions: paid plans
MVP: about 100 users, $1.80/mo
Someone with an idea or a solo founder. Built with an AI app builder: describe it and it builds and hosts it.
| What | Pick | Monthly |
|---|---|---|
| AI coding agent | Base44 Starter | $16 |
| Mailgun Free | $0 | |
| Payments | Stripe Standard (US cards) | $17.50 |
| Analytics & errors | PostHog Free | $0 |
| AI API | Anthropic (Claude API) Claude Haiku 5.5 | $1.80 |
Growth: about 10,000 users, $223.94/mo
A funded team or one with traction. Built with an AI coding agent for each of about 3 developers.
- Error tracking: know when something breaks
- Daily backups: daily backups once real people depend on it
| What | Pick | Monthly |
|---|---|---|
| AI coding agent | Augment Code Standard | $20 |
| Hosting | Fly.io Machine, shared-cpu-1x with 256MB RAM | $6.19 |
| Database | DigitalOcean Managed PostgreSQL Single node (1 vCPU, 2 GiB RAM, 30 GiB disk) | $34.75 |
| Sign-in | Auth0 Free | $0 |
| Amazon SES A la carte | $3 | |
| Payments | Stripe Standard (US cards) | $1,750 |
| Analytics & errors | PostHog Pay as you go | $0 |
| AI API | Anthropic (Claude API) Claude Haiku 5.5 | $180 |
First cost jump: DigitalOcean Managed PostgreSQL: Single node (1 vCPU, 2 GiB RAM, 30 GiB disk) limits reached, upgrade to Single node (2 vCPU, 4 GiB RAM, 60 GiB disk)
Scale: about 100,000 users, $2,126.19/mo
An established product. Built with an AI coding agent with team billing and admin for about 10 developers.
- Error tracking: know when something breaks
- Daily backups: daily backups
- Point-in-time recovery: restore to any minute after a bad deploy
- Multi-factor authentication: two-factor sign-in for accounts worth stealing
- SOC 2 report available: customers' security reviews ask for a SOC 2 report
| What | Pick | Monthly |
|---|---|---|
| AI coding agent | Augment Code Standard | $20 |
| Hosting | Fly.io Machine, shared-cpu-1x with 256MB RAM | $42.19 |
| Database | Neon Scale | $225 |
| Sign-in | WorkOS AuthKit | $0 |
| Amazon SES A la carte | $30 | |
| Payments | Stripe Standard (US cards) | $17,500 |
| Analytics & errors | Sentry Team | $29 |
| AI API | Anthropic (Claude API) Claude Haiku 5.5 | $1,800 |
Assumed usage: Each person asks about 30 questions a month, each with about 4,000 tokens of passages in and 400 out; embeddings take about 5 MB per person.
What drives the cost
- Model tokens per question, plus embedding every document once
- Database storage for embeddings
- Re-embedding when you change models
Lock-in
Embeddings are tied to the model that made them; switching models means re-embedding everything.
Use this in Claude (with Kong Reviews connected at https://kongreviews.com/mcp)

