😂 Joke Generator

GPT OSS 120B via OpenAI API

Concept

This demo asks an LLM to tell a joke about a topic you supply. It uses the OpenAI API as the inference provider to hit an open-source model endpoint. Note: the python code uses historical variable names like get_groq_client.

The request uses a high temperature (1.3) so the model leans toward creative, varied output instead of repeating the same few jokes it defaults to at lower temperatures.

Theory & Concepts

Unleashing Creativity with Temperature and Persona

This Joke Generator demo relies on two key concepts to produce varied and funny responses:

  • High Temperature for Variety: Language models predict the next word based on probabilities. By default, they pick the most likely word. In this demo, we intentionally set the temperature to a high value of 1.3 in the API call. This forces the model to take creative risks and pick less obvious words, preventing it from repeating the exact same jokes when you ask for a joke about "dogs" multiple times.
  • Role-Based Prompting: We use the system role to define the model's persona, telling it: "You are a comedian with a huge repertoire of jokes. Never repeat the same joke twice." This anchors its behavior. Then, we pass your specific topic in the user prompt: "Tell me a joke about {subject}. Just the joke, no commentary." This combination ensures you get exactly what you asked for—a raw joke without unnecessary introductory text.

Request flow

🧑 Browser Type a topic, click “Tell me a joke”
→
POST /joke Flask route
→
get_joke() Builds the prompt
→
OpenAI API GPT OSS 120B generates the joke
→
🧑 Browser Joke rendered in the result box

Code flow

flowchart TD A[Browser
topic] -->|POST /joke| B[app.py
joke route] B -->|topic| C[joke.py
get_joke] C -->|prompt + temperature| D[OpenAI API
GPT OSS 120B] D -->|joke text| C C -->|joke text| B B -->|JSON result| A

Backend

Generates the joke via OpenAI API's chat completions API. 
from config import get_groq_client

# GPT OSS 120B via OpenAI API: fast inference.
JOKE_MODEL = "openai/gpt-oss-120b"


def get_joke(topic: str = "") -> str:
    # Fall back to "random" instead of sending an empty string to the model.
    subject = topic.strip() if topic.strip() else "random"

    client = get_groq_client()
    response = client.chat.completions.create(
        model=JOKE_MODEL,
        temperature=1.3,  # high temperature = more creative, varied jokes
        messages=[
            {
                "role": "system",
                # Persona + "never repeat" nudge the model toward variety.
                "content": "You are a comedian with a huge repertoire of jokes. Never repeat the same joke twice.",
            },
            {
                "role": "user",
                # Ask for just the joke so the UI can display it directly.
                "content": f"Tell me a joke about {subject}. Just the joke, no commentary.",
            },
        ],
    )
    # choices[0] is the first (and only) completion; .message.content is the
    # assistant's reply text.
    return response.choices[0].message.content

API route

Exposes the joke generator over HTTP as POST /joke. 
@bp.route("/joke", methods=["POST"])
def joke():
    data = request.get_json(force=True)
    # topic is optional; an empty string makes get_joke() pick randomly.
    topic = (data.get("topic") or "").strip()
    try:
        text = get_joke(topic)
        response = jsonify({"result": text})
        # Every click should fetch a fresh joke — never serve a cached one.
        response.headers["Cache-Control"] = "no-store"
        return response
    except Exception as e:
        return jsonify({"error": str(e)}), 500