Transparent GPT

Educational DemoAI Model: checking

System Prompt

Usually Hidden

System prompts are typically hidden from users in production...

LLM

Transparent LLM Demo

Type a message to see how it flows through the system. Watch the pipeline animate step by step, then see the response type out. You have 3 turns per conversation to ask questions.

0/300

Turn 0 of 3

Data Flow Pipeline

Send a message to begin the visualization

Client to Server

Raw text string:

"Hello, how are you?"

Wrapped with role:

{
  "role": "user",
  "content": "Hello, how are you?"
}

Full API payload (Groq):

{
  "model": "llama-3.1-8b-instant",
  "messages": [
    {
      "role": "system",
      "content": "You are TransparentGPT, an interactive AI chat usi..."
    }
  ],
  "temperature": 0.7,
  "max_tokens": 200
}

Inside Groq Server

Server validates:

✓API key valid

✓Rate limit OK

✓Model available

✓Content policy passed

Text to Token IDs (BPE algorithm):

Hello[9906],[11] how[1268] are[527] you[499]?[30]

Token IDs to 12,288-dimensional vectors:

Token "Hello" (ID: 9906)

[0.023, -0.156, 0.892, 0.445, -0.234, 0.667, ...]

+ positional encoding added for sequence position

96 layers with 96 attention heads each:

Q x K^T

Query-Key dot product

Softmax

Normalize attention

x Values

Weighted combination

FFN

Feed-forward network

Softmax over 100K+ vocabulary (llama-3.1 tokenizer):

"I'm"

28.3%

"Hello"

22.1%

"Hey"

15.6%

⚡ Simulated probabilities based on input pattern • temperature=0.7

Server to Client

Output token IDs back to text:

Hello, how are you?

Streaming via Server-Sent Events:

data: {"choices":[{"delta":{"content":"Response"}}]}

data: {"choices":[{"delta":{"content":" in progress"}}]}

...

data: [DONE]