LLM Integration

Prerequisites:

• npm install @jaypie/llm
• API key for at least one provider (Anthropic, Fireworks, Google, Meta, Mistral, OpenAI, OpenRouter, or xAI)

Overview

@jaypie/llm provides a unified interface for calling multiple LLM providers. The Llm class handles provider-specific implementations while exposing a consistent API for operate, stream, and tool calling.

Quick Reference

Supported Providers

Models are listed as LLM.MODEL.* names rather than ids. The alias is stable across releases; the id behind it moves whenever the provider ships a successor.

Provider	Models	Env Variable
Anthropic	MODEL.SONNET, MODEL.OPUS, MODEL.HAIKU, MODEL.FABLE	ANTHROPIC_API_KEY
Bedrock	MODEL.NOVA_PRO, MODEL.NOVA_LITE	(AWS credentials)
Fireworks	MODEL.FIREWORKS.* (DEEPSEEK, GLM, GPT_OSS, INKLING, KIMI, MINIMAX, NEMOTRON, QWEN)	FIREWORKS_API_KEY
Google	MODEL.GEMINI_FLASH, MODEL.GEMINI_FLASH_LITE, MODEL.GEMINI_PRO	GOOGLE_API_KEY
Meta	MODEL.MUSE_SPARK, MODEL.MUSE_SPARK_CONTRIBUTOR	META_API_KEY (or MODEL_API_KEY)
Mistral	MODEL.MISTRAL.* (LARGE, MEDIUM, SMALL, OCR)	MISTRAL_API_KEY
OpenAI	MODEL.ASTRA, MODEL.SOL, MODEL.LUNA, MODEL.TERRA	OPENAI_API_KEY
OpenRouter	MODEL.OPENROUTER.* (GLM, LUNA, SONNET)	OPENROUTER_API_KEY
xAI	MODEL.GROK	XAI_API_KEY

Core Methods

Method	Purpose	Returns
Llm.operate()	Single prompt/response	string
Llm.stream()	Streaming response	AsyncGenerator
llm.operate()	Instance method with history	string

Basic Usage

Static Method (Recommended for single calls)

import Llm, { LLM } from "@jaypie/llm";

// Provider auto-detected from model
const response = await Llm.operate("What is 2+2?", {
  model: LLM.MODEL.SONNET,
});
console.log(response); // "4"

Instance Method (For conversations)

import Llm, { LLM } from "@jaypie/llm";

const llm = new Llm("anthropic", {
  model: LLM.MODEL.SONNET,
  system: "You are a helpful assistant.",
});

const response1 = await llm.operate("My name is Alice.");
const response2 = await llm.operate("What is my name?");
// Maintains conversation history

The first constructor argument may be a provider name or a model name. A model name auto-detects the provider and is retained:

const llm = new Llm("claude-sonnet-4-6"); // -> anthropic, claude-sonnet-4-6

Model Selection

By Model Name

// Anthropic models
await Llm.operate(prompt, { model: LLM.MODEL.SONNET });
await Llm.operate(prompt, { model: LLM.MODEL.OPUS });
await Llm.operate(prompt, { model: LLM.MODEL.HAIKU });

// OpenAI models
await Llm.operate(prompt, { model: LLM.MODEL.SOL });
await Llm.operate(prompt, { model: LLM.MODEL.LUNA });

// Google models
await Llm.operate(prompt, { model: LLM.MODEL.GEMINI_FLASH });

Explicit Provider

const llm = new Llm("openai", {
  model: LLM.MODEL.SOL,
});

Streaming

Basic Streaming

for await (const chunk of Llm.stream("Tell me a story")) {
  process.stdout.write(chunk.content || "");
}

Express SSE Streaming

import { expressStreamHandler, createExpressStream } from "jaypie";
import Llm from "@jaypie/llm";

export default expressStreamHandler(async (req, res, context) => {
  const stream = createExpressStream(context);

  for await (const chunk of Llm.stream(req.body.prompt)) {
    stream.write(chunk.content || "");
  }

  stream.end();
});

Lambda Response Streaming

import { lambdaStreamHandler } from "jaypie";
import Llm from "@jaypie/llm";

export const handler = awslambda.streamifyResponse(
  lambdaStreamHandler(
    async (event, context) => {
      for await (const chunk of Llm.stream(event.prompt)) {
        context.responseStream.write(`data: ${JSON.stringify(chunk)}\n\n`);
      }
    },
    {
      contentType: "text/event-stream",
    },
  ),
);

Tool Calling

Creating Tools

import Llm, { LLM, Toolkit } from "@jaypie/llm";

const toolkit = new Toolkit([
  {
    name: "get_weather",
    description: "Get current weather for a city",
    parameters: {
      type: "object",
      properties: {
        city: {
          type: "string",
          description: "City name",
        },
      },
      required: ["city"],
    },
    call: async ({ city }) => {
      // Actual implementation
      return `Weather in ${city}: sunny, 72F`;
    },
  },
]);

const response = await Llm.operate("What's the weather in NYC?", {
  model: LLM.MODEL.SONNET,
  tools: toolkit,
});

Zod Schema Tools

import { z } from "zod";
import Llm, { Toolkit } from "@jaypie/llm";

const toolkit = new Toolkit([
  {
    name: "create_user",
    description: "Create a new user",
    parameters: z.object({
      name: z.string().describe("User's full name"),
      email: z.string().email().describe("User's email"),
      age: z.number().optional().describe("User's age"),
    }),
    call: async ({ name, email, age }) => {
      const user = await db.users.create({ name, email, age });
      return { id: user.id, name, email };
    },
  },
]);

Tool with Validation

{
  name: "transfer_money",
  description: "Transfer money between accounts",
  parameters: {
    type: "object",
    properties: {
      from: { type: "string" },
      to: { type: "string" },
      amount: { type: "number" },
    },
    required: ["from", "to", "amount"],
  },
  call: async ({ from, to, amount }) => {
    if (amount <= 0) {
      throw BadRequestError("Amount must be positive");
    }
    // Process transfer
    return { success: true, transactionId: uuid() };
  },
}

Structured Output

Pass format to receive guaranteed-valid JSON. format accepts Jaypie’s natural schema syntax (preferred), a raw JSON Schema, or a Zod schema.

Natural Schema

const response = await Llm.operate(
  "Extract the person's name and age from: 'John is 25 years old'",
  {
    model: LLM.MODEL.SOL,
    format: {
      name: String,
      age: Number,
    },
  },
);
// Returns: { name: "John", age: 25 }

Every array field declared in format is guaranteed present in the response as an array — an empty result surfaces as [], never undefined.

JSON Schema

const response = await Llm.operate(prompt, {
  model: LLM.MODEL.SOL,
  format: {
    type: "object",
    properties: {
      name: { type: "string" },
      age: { type: "number" },
    },
  },
});
// Returns: { name: "John", age: 25 }

Zod Schema

import { z } from "zod";

const PersonSchema = z.object({
  name: z.string(),
  age: z.number(),
});

const response = await Llm.operate(prompt, {
  model: LLM.MODEL.SOL,
  format: PersonSchema,
});
// Returns typed object matching PersonSchema

Multi-Turn Conversations

const llm = new Llm("anthropic", {
  model: LLM.MODEL.SONNET,
  system: "You are a math tutor.",
});

// Conversation history maintained automatically
await llm.operate("What is calculus?");
await llm.operate("Can you give me an example?");
await llm.operate("How is that used in physics?");

// Access history
console.log(llm.history);

Files and Images

Image Input

const response = await Llm.operate("What's in this image?", {
  model: LLM.MODEL.SONNET,
  files: [
    {
      type: "image",
      data: base64ImageData,
      mediaType: "image/png",
    },
  ],
});

URL Image

const response = await Llm.operate("Describe this image", {
  model: LLM.MODEL.SOL,
  files: [
    {
      type: "image",
      url: "https://example.com/image.jpg",
    },
  ],
});

Hooks

Lifecycle Hooks

Seven hooks fire through the operate loop, each receiving the full provider request/response payloads:

const response = await Llm.operate(prompt, {
  model: LLM.MODEL.SONNET,
  hooks: {
    beforeEachModelRequest: ({ providerRequest }) => {
      log.trace("[llm] calling model");
    },
    afterEachModelResponse: ({ content, usage }) => {
      log.trace("[llm] response received");
      log.var({ tokens: usage[usage.length - 1]?.total });
    },
    beforeEachTool: ({ toolName, args, message }) => {
      // message is the tool's resolved LlmTool.message, when defined
      log.trace(message ?? `[llm] calling ${toolName}`);
    },
    afterEachTool: ({ result, toolName, message }) => {
      log.trace(`[llm] ${toolName} returned`);
    },
    onToolError: ({ error, toolName, message }) => {
      log.warn(`[llm] ${toolName} failed`);
      log.var({ error: error.message });
    },
    onRetryableModelError: ({ error }) => {
      log.warn("[llm] model call failed, retrying");
    },
    onUnrecoverableModelError: ({ error }) => {
      log.error("[llm] model call failed");
      log.var({ error });
    },
  },
});

Progress Events

For progress reporting (UI updates, websockets, queue notifications), prefer a single onProgress callback over wiring individual hooks. It receives lightweight, serializable events as the operate loop runs:

const response = await Llm.operate(prompt, {
  model: LLM.MODEL.SONNET,
  tools: toolkit,
  onProgress: (event) => {
    // event.type: start, model_request, model_response,
    //             tool_call, tool_result, tool_error, retry, done
    websocket.send(JSON.stringify(event));
  },
});

Fields carried by each event (turn is 1-indexed):

Event	Fields
start	model, provider, maxTurns
model_request	turn, model
model_response	turn, content (text, if any), toolCalls ([{ name, arguments }], if any), usage (this turn)
tool_call	turn, tool: { name, arguments, message } — fires before the tool runs; arguments is the JSON string; message is the resolved LlmTool.message, when the tool defines one
tool_result	turn, tool: { name } — the result value is deliberately omitted (it can be arbitrarily large); use the afterEachTool hook to receive it
tool_error	turn, tool: { name }, error (message string)
retry	turn, error (message string)
done	turn (total turns used), content (final text or structured output), usage (cumulative)

Errors thrown by the callback are logged and never interrupt the loop. stream() communicates progress through its chunks; onProgress applies to operate().

Error Handling

import { BadGatewayError, log } from "jaypie";
import Llm, { LLM } from "@jaypie/llm";

async function askLlm(prompt) {
  try {
    return await Llm.operate(prompt, { model: LLM.MODEL.SONNET });
  } catch (error) {
    log.error("[askLlm] LLM call failed");
    log.var({ error: error.message });
    throw BadGatewayError("AI service unavailable");
  }
}

Provider-Specific Options

modelOptions sets provider-specific request fields per model. Keys are an exact model id, a MODEL catalog key, a provider (or the alias gemini), or default, and every attempt in a fallback chain merges the keys that match its own model. providerOptions is deprecated, reaches the primary only, and is removed in 2.0.

Anthropic

max_tokens defaults to the model’s maximum output (capped at 16,384 for non-streaming requests). Override it through modelOptions:

await Llm.operate(prompt, {
  model: LLM.MODEL.SONNET,
  modelOptions: { anthropic: { max_tokens: 32000 } },
  temperature: 0.7,
});

OpenAI

await Llm.operate(prompt, {
  model: LLM.MODEL.SOL,
  temperature: 0.5,
  topP: 0.9,
});

Related

• Express on Lambda - Building APIs with LLM
• Handler Lifecycle - Integrating LLM in handlers
• @jaypie/llm - Full API reference
• Fabric - Service handler to LLM tool conversion