Starter templates

npm create rag-pipeline scaffolds a project that answers questions from a folder of documents, cites the passages it used and checks every sentence of the answer against them. Each starter runs before you add an API key, returning the matching passages, so you can try retrieval first.

npm create rag-pipeline@latest my-app -- --template next
cd my-app
npm install
npm run dev

Without --template, it asks which starter to use.

Template	What you get	Run it
next	A Next.js app: a question page and POST /api/ask, with any AI SDK model wrapped in groundingMiddleware (retrieval before the call, a grounding check after)	npm run dev
mastra	A Mastra agent with a search_documents tool, a groundedness scorer (no model calls) and guardrails (prompt injection in, PII redaction out)	npm run ask -- "..."
workers	A Cloudflare Worker: GET /ask as JSON or as a token stream, with citations and a groundedness score, on the library's /edge build	npm run dev

All three index the sample documents in memory with the built-in TF-IDF embedder, so there is no database and no embedding API to set up. Replace the documents (docs/ in next and mastra, src/docs.js in workers) with your own.

Add a model

Copy the example file and set ANTHROPIC_API_KEY:

Template	Copy	To	Model setting
next	.env.example	.env.local	ANTHROPIC_MODEL, or any AI SDK provider
mastra	.env.example	.env	MODEL, any Mastra model (anthropic/...)
workers	.dev.vars.example	.dev.vars	ANTHROPIC_MODEL; ANTHROPIC_BASE_URL for AI Gateway

The example files are copied, the files with keys are in .gitignore.

What the answer looks like

From the next starter, with a key:

{
  "answer": "Full-time employees receive twenty days of paid vacation per year. Unused days carry over until the end of March.",
  "sources": [{ "id": "vacation", "source": "vacation.md" }],
  "grounding": { "grounded": true, "groundedness": 1, "unsupported": [] }
}

When a sentence isn't supported by any passage, grounded is false and unsupported lists it. The check is lexical (see Evaluation): a correct paraphrase can be reported unsupported, and the optional LLM judge can rule on those sentences.

Next steps

• More documents: replace MemoryRetriever with QdrantRetriever or PgVectorRetriever, and LocalEmbedder with an embedding model (Connectors).
• Test the answers in CI: rag-pipeline eval scores a test set and fails the build when answers regress (CLI).
• Deploy: Vercel for next (the documents are included with the route that reads them), node .mastra/output/index.mjs or a Mastra deployer for mastra, npm run deploy for workers.

How the starters are tested

Every change to the library runs npm run test:templates in CI: each starter is scaffolded, installed with that build of the library, built (next build, tsc and mastra build, wrangler deploy --dry-run) and asked a question without a key. Before a release, the same script runs with a Claude key and checks that each answer is grounded.