Agent Skills: Groq Events & Async Patterns

Groq Events & Async Patterns

Overview

Build event-driven architectures around Groq's inference API. Groq does not provide native webhooks, but its sub-second latency enables unique patterns: real-time SSE streaming, batch processing with callbacks, queue-based pipelines, and event processors that use Groq as an LLM classification/extraction engine.

This skill uses Read, Write, and Edit to scaffold and update these handlers in your codebase, and curl to exercise the resulting endpoints. Step 1 (the SSE endpoint) is inline below; the batch, webhook-processor, health-monitor, and Python async patterns live in references/implementation.md.

Prerequisites

• groq-sdk (Node) or groq (Python) installed, GROQ_API_KEY set
• Queue system for batch patterns (BullMQ, Redis, SQS)
• Understanding of Server-Sent Events (SSE) for streaming

Authentication

Groq authenticates with a single API key. Export GROQ_API_KEY in the environment and the SDK reads it automatically — never hard-code the key or embed it in a request body. The key is a bearer credential; treat it like any secret (env var or secrets manager, never committed). No per-request auth headers are needed when the SDK is constructed with new Groq() / AsyncGroq().

Instructions

Write each handler as a file in your project (Read/Write/Edit), then drive it with curl to confirm behavior.

Step 1: SSE Streaming Endpoint

Stream tokens to the browser as they are generated. Set the text/event-stream headers, disable proxy buffering with X-Accel-Buffering: no, and write one data: frame per token, ending with a done event.

import Groq from "groq-sdk";
import express from "express";

const groq = new Groq();
const app = express();
app.use(express.json());

app.post("/api/chat/stream", async (req, res) => {
  const { messages, model = "llama-3.3-70b-versatile" } = req.body;

  res.writeHead(200, {
    "Content-Type": "text/event-stream",
    "Cache-Control": "no-cache",
    Connection: "keep-alive",
    "X-Accel-Buffering": "no",  // Disable nginx buffering
  });

  try {
    const stream = await groq.chat.completions.create({
      model,
      messages,
      stream: true,
      max_tokens: 2048,
    });

    for await (const chunk of stream) {
      const content = chunk.choices[0]?.delta?.content;
      if (content) {
        res.write(`data: ${JSON.stringify({ content, type: "token" })}\n\n`);
      }
    }

    res.write(`data: ${JSON.stringify({ type: "done" })}\n\n`);
  } catch (err: any) {
    res.write(`data: ${JSON.stringify({ type: "error", message: err.message })}\n\n`);
  }

  res.end();
});

Steps 2–5: Batch, Webhook Processor, Health Monitor, Python Async

The remaining patterns follow the same shape — Groq as a fast inference engine behind a queue or an event loop. Each is documented in full, with runnable code, in references/implementation.md:

• Step 2 — Batch processing with BullMQ: enqueue prompts, process with a rate-limited worker (concurrency: 5, limiter: 25 RPM), fire a callback per item.
• Step 3 — Webhook event processor: ack the sender with 202 immediately, then classify/extract the event asynchronously with llama-3.1-8b-instant.
• Step 4 — Scheduled health monitor: ping each model with a one-token request on an interval, tracking latency and tokens/sec.
• Step 5 — Python async batch: asyncio.Semaphore + gather for concurrent processing without a queue.

Output

Each pattern produces a distinct, observable artifact you can assert against:

• SSE endpoint — a text/event-stream response: one data: {"content":…,"type":"token"} frame per token, terminated by data: {"type":"done"} (or a type:"error" frame on failure).
• Batch worker — a groq.batch.item_completed callback POST per prompt, carrying batchId, index, total, content, model, and token usage.
• Webhook processor — an immediate 202 {"received": true} ack, followed by a background classification object {type, priority, summary, action}.
• Health monitor — a per-model record {status, latencyMs, tokensPerSec} (or {status:"error", error}) logged each interval.

See references/examples.md for the concrete payloads.

Event Pattern Summary

| Pattern | Groq Model | Latency | Use Case | |---------|-----------|---------|----------| | SSE streaming | llama-3.3-70b-versatile | ~200ms TTFT | Real-time chat | | Batch queue | llama-3.1-8b-instant | ~80ms TTFT | Document processing | | Webhook processor | llama-3.1-8b-instant | ~80ms TTFT | Event classification | | Health monitor | llama-3.1-8b-instant | ~80ms TTFT | Uptime tracking |

Error Handling

| Issue | Cause | Solution | |-------|-------|----------| | SSE disconnect | Client timeout or network | Implement reconnection with last-event-id | | Batch item fails | Rate limit or model error | Queue retry with exponential backoff | | Webhook timeout | Processing takes too long | Acknowledge immediately (202), process async | | Health check 429 | Monitoring consuming quota | Reduce check frequency, use smallest model |

Examples

Worked, runnable examples — consuming the SSE endpoint with curl, submitting a batch and receiving callbacks, and classifying an inbound webhook — are in references/examples.md. A minimal first call:

curl -N -X POST http://localhost:3000/api/chat/stream \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"Explain SSE in one sentence."}]}'

Resources

• Full implementation walkthrough — Steps 2–5 with runnable code
• Worked examples — curl calls and expected payloads
• Groq API Reference
• Groq Text Generation (streaming)
• BullMQ Documentation

For performance optimization, see the groq-performance-tuning skill.