RAG PipelineAI Workflow

“Run `npx radzor@latest add embeddings-store llm-completion` to install 2 Radzor components. Then read components/radzor/embeddings-store/radzor.manifest.json, components/radzor/llm-completion/radzor.manifest.json and each component's llm/integration.md. Wire them together to plug-and-play Retrieval-Augmented Generation: embed text chunks in-memory, search by semantic similarity, and generate a context-aware response with an LLM. Use the manifest's inputs (check envVar for required environment variables), outputs (check fields for object shapes), composability (check mapField for field extraction), and actions — don't invent custom interfaces.”

Paste this into Claude Code, Cursor, Windsurf, or any AI coding agent.

// npx radzor@latest add embeddings-store llm-completion
import { EmbeddingsStore } from "./components/radzor/embeddings-store"
import { LLMCompletion }   from "./components/radzor/llm-completion"

const store = new EmbeddingsStore({
  provider: "openai",
  apiKey: process.env.OPENAI_KEY!,
  model: "text-embedding-3-small",
})
const llm = new LLMCompletion({
  provider: "openai",
  apiKey: process.env.OPENAI_KEY!,
  model: "gpt-4o",
  maxTokens: 1024,
})

// --- 1. Ingestion: split your docs into chunks and add them ---
const chunks = [
  "Radzor components are installed with npx radzor@latest add.",
  "Each component has a radzor.manifest.json describing its API.",
  "Components can be wired together using their inputs and outputs.",
]

for (const [i, text] of chunks.entries()) {
  await store.add(`chunk-${i}`, text)
}

// --- 2. Query: search for relevant chunks, then ask the LLM ---
const userQuery = "How do I install a component?"
const results = await store.search(userQuery, 3)
const context = results.map((r) => r.text).join("\n")

const { content } = await llm.complete(
  `Answer based on the following context only.\n\nContext:\n${context}\n\nQuestion: ${userQuery}`
)
console.log(content)