Starter stack

Chat with your docs (RAG)

Your product answers questions from a pile of documents — help docs, contracts, a knowledge base. You need to ingest files, search them by meaning, and stream answers that cite their sources.

34 makers' products and 74 open-source projects use at least half of these 8 tools.

What it costs

Launchabout 1,000 users: $26/mo
Tractionabout 10,000 users: $104/mo
Scaleabout 100,000 users: $447/mo

Monthly bills of the priced picks, list prices.

Do the picks fit together?No conflicts: no acquired or archived tool, no two tools for a decision that usually has one, nothing the plan's rules say doesn't fit.Open in Check my stack — add or swap tools →

Hand it to your AI agent

Paste into AGENTS.md, CLAUDE.md or your Cursor rules: each tool with its agent docs, MCP server and CLI where they exist.

Preview

# Stack: Chat with your docs (RAG)

Your product answers questions from a pile of documents — help docs, contracts, a knowledge base. You need to ingest files, search them by meaning, and stream answers that cite their sources.

Use these tools. Ask before swapping one for an alternative.

- Frontend Framework: Next.js — https://nextjs.org
agent docs: https://nextjs.org/llms.txt · CLI: npx next
Why: Streaming chat UI and the API routes that call the model, in one app.
- Database: Supabase — https://supabase.com
agent docs: https://supabase.com/llms.txt · MCP server: com.supabase/mcp (https://mcp.supabase.com/mcp) · CLI: npx supabase
Why: Postgres for users and documents, file storage for the uploads, and the pgvector extension one click away.
- Vector Database: pgvector — https://github.com/pgvector/pgvector
agent docs: https://github.com/llms.txt
Why: Embeddings live next to the rows they describe, so filtering by user or workspace is a normal SQL WHERE — no second database to keep in sync.
- AI SDK & Agent Framework: AI SDK — https://ai-sdk.dev
Why: Embedding calls, streaming answers and tool calls with one API, and a provider switch when a cheaper model is good enough.
- LLM API: OpenAI API — https://openai.com
agent docs: https://developers.openai.com/llms.txt
Why: Embedding and chat models from one account; the most examples to copy for retrieval.
- Background Jobs & Cron: Trigger.dev — https://trigger.dev
agent docs: https://trigger.dev/llms.txt
Why: Parsing and embedding a 300-page PDF takes minutes — run it as a job with retries, not inside a request.
- LLM Observability & Evals: Langfuse — https://langfuse.com
agent docs: https://langfuse.com/llms.txt · CLI: npx langfuse-cli
Why: See which chunks each answer retrieved, what it cost, and whether a prompt change made answers better — the difference between a demo and a product.
- Hosting: Vercel — https://vercel.com
agent docs: https://vercel.com/llms.txt · MCP server: com.vercel/vercel-mcp (https://mcp.vercel.com) · CLI: npx vercel
Why: Streaming responses and preview deploys out of the box.

Trade-offs: pgvector is plenty until millions of vectors or heavy hybrid search; then a dedicated store like Qdrant or turbopuffer earns its keep. Retrieval quality depends far more on how you chunk and label documents than on any tool here — budget time for evals.

Source: https://howmakersbuild.com/stacks/rag-app

The picks

Frontend Framework

Next.jsllms.txt CLI

Streaming chat UI and the API routes that call the model, in one app.

Chosen for frontend framework by 1510 makers' products we track · the alternatives →

Database

Supabasellms.txt MCP CLIIn Bolt, Lovable, v0

Postgres for users and documents, file storage for the uploads, and the pgvector extension one click away.

Chosen for database by 564 makers' products we track · the alternatives →

Vector Database

pgvectorllms.txt

Embeddings live next to the rows they describe, so filtering by user or workspace is a normal SQL WHERE — no second database to keep in sync.

No maker's product we track shows it for vector database yet · the alternatives →

AI SDK & Agent Framework

AI SDK

Embedding calls, streaming answers and tool calls with one API, and a provider switch when a cheaper model is good enough.

Chosen for AI SDK & agent framework by 4 makers' products we track · the alternatives →

LLM API

OpenAI APIllms.txtIn Lovable, Replit

Embedding and chat models from one account; the most examples to copy for retrieval.

Chosen for LLM API by 734 makers' products we track · the alternatives →

Background Jobs & Cron

Trigger.devllms.txt

Parsing and embedding a 300-page PDF takes minutes — run it as a job with retries, not inside a request.

Chosen for background jobs & cron by 23 makers' products we track · the alternatives →

LLM Observability & Evals

Langfusellms.txt CLI

See which chunks each answer retrieved, what it cost, and whether a prompt change made answers better — the difference between a demo and a product.

Chosen for LLM observability & evals by 33 makers' products we track · the alternatives →

Hosting

Vercelllms.txt MCP CLIIn v0

Streaming responses and preview deploys out of the box.

No maker's product we track shows it for hosting yet · the alternatives →

What this costs you

pgvector is plenty until millions of vectors or heavy hybrid search; then a dedicated store like Qdrant or turbopuffer earns its keep. Retrieval quality depends far more on how you chunk and label documents than on any tool here — budget time for evals.

Ask your AI about this, with this page as the source:ChatGPT ↗Claude ↗Perplexity ↗

Who builds this way

Using at least half of these tools, each with a source: makers' products from their own words, customer stories, subprocessor lists or websites; open-source projects from their code.