Cria is a lightweight prompt composition library for structured prompt engineering. Build prompts as components, keep behavior predictable, and reuse the same structure across providers. Runs on Node, Deno, Bun, and Edge; adapters require their SDKs.

const messages = await cria
  .prompt()
  .system("You are a research assistant.")
  .vectorSearch({ store, query: question, limit: 10 })
  .provider(openai("gpt-5-nano"), p => p
    .summary(conversation, { store: memory })
    .last(conversation, { N: 20 })
  )
  .user(question)
  .render({ budget: 200_000, renderer });

Start with Quickstart, then use Docs to jump to the right how-to.

Use Cria when you need...

• Need RAG? Call .vectorSearch({ store, query }).
• Need a summary for long conversations? Use .summary(...).
• Need to cap history but keep structure? Use .last(...).
• Need to drop optional context when the context window is full? Use .omit(...).
• Using AI SDK? Plug and play with @fastpaca/cria/ai-sdk!
• Prefer TSX? Import the optional JSX surface from @fastpaca/cria/jsx.

Integrations

import OpenAI from "openai";
import { chatCompletions } from "@fastpaca/cria/openai";
import { cria } from "@fastpaca/cria";

const client = new OpenAI();
const messages = await cria
  .prompt()
  .system("You are helpful.")
  .user(userQuestion)
  .render({ budget, tokenizer, renderer: chatCompletions });
const response = await client.chat.completions.create({ model: "gpt-4o-mini", messages });

import OpenAI from "openai";
import { responses } from "@fastpaca/cria/openai";
import { cria } from "@fastpaca/cria";

const client = new OpenAI();
const input = await cria
  .prompt()
  .system("You are helpful.")
  .user(userQuestion)
  .render({ budget, tokenizer, renderer: responses });
const response = await client.responses.create({ model: "gpt-5-nano", input });

import Anthropic from "@anthropic-ai/sdk";
import { anthropic } from "@fastpaca/cria/anthropic";
import { cria } from "@fastpaca/cria";

const client = new Anthropic();
const { system, messages } = await cria
  .prompt()
  .system("You are helpful.")
  .user(userQuestion)
  .render({ budget, tokenizer, renderer: anthropic });
const response = await client.messages.create({ model: "claude-haiku-4-5", system, messages });

import { renderer } from "@fastpaca/cria/ai-sdk";
import { cria } from "@fastpaca/cria";
import { generateText } from "ai";

const messages = await cria
  .prompt()
  .system("You are helpful.")
  .user(userQuestion)
  .render({ budget, tokenizer, renderer });
const { text } = await generateText({ model, messages });

Roadmap

Done

• [x] Fluent DSL and priority-based eviction
• [x] Components: Region, Message, Truncate, Omit, Last, Summary, VectorSearch, ToolCall, ToolResult, Reasoning, Examples, CodeBlock, Separator
• [x] Renderers: OpenAI (Chat Completions + Responses), Anthropic, AI SDK
• [x] AI SDK helpers: Messages component, DEFAULT_PRIORITIES
• [x] Memory: InMemoryStore, Redis, Postgres, Chroma, Qdrant
• [x] Observability: render hooks, validation schemas, snapshots, OpenTelemetry
• [x] Tokenizer helpers

Planned

• [ ] Next.js adapter
• [ ] GenAI semantic conventions for OpenTelemetry
• [ ] Visualization tool
• [ ] Prompt eval / testing functionality

Contributing

• Issues and PRs are welcome.
• Keep changes small and focused.
• If you add a feature, include a short example or doc note.
• Prefer DSL-based examples in contributions; JSX lives under @fastpaca/cria/jsx as optional sugar.

Support

• Open a GitHub issue for bugs or feature requests.
• For quick questions, include a minimal repro or snippet.

FAQ

• Does this replace my LLM SDK? No - Cria builds prompt structures. You still use your SDK to call the model.
• How do I tune token budgets? Pass budget to render() and set priorities on regions; see docs/how-to/fit-and-compaction.md.
• Is this production-ready? Not yet! It is a work in progress and you should test it out before you run this in production.

License

MIT