LLM Integration Pipeline

LLM Integration Pipeline

Udility Diffuser leverages a multi-stage LLM pipeline to bridge the gap between abstract natural language descriptions and precise SVG (Scalable Vector Graphics) code. By utilizing high-parameter models via OpenRouter, the system ensures that generated illustrations are both syntactically correct and educationally relevant.

Core Model & Provider

The pipeline is powered by the Hermes-3-Llama-3.1-405B model, hosted via OpenRouter. This model was selected for its superior reasoning capabilities and its ability to follow complex structural instructions, which is critical for generating valid SVG markup.

• Provider: OpenRouter
• Model String: nousresearch/hermes-3-llama-3.1-405b
• Primary Engine: Meta Llama 3.1 (405B parameter version)

Orchestration Workflow

The integration follows a sequential two-step inference process to maximize the detail and accuracy of the output:

1. Instruction Synthesis: The user's prompt is first sent to the LLM to generate a "visual blueprint." This stage expands a simple prompt (e.g., "Amoeba lifecycle") into a detailed set of technical instructions for drawing specific shapes, labels, and coordinates.
2. SVG Code Generation: The visual blueprint is then fed back into the LLM with a specialized system prompt. This step forces the model to act as a code generator, outputting raw <svg> code based on the synthesized instructions.

Public Interface

Users interact with the pipeline primarily through the generate_image_from_text function. This high-level wrapper manages the API calls to OpenRouter, processes the SVG output, and handles the conversion to a viewable image format.

generate_image_from_text

Generates an illustrative PNG image from a text-based prompt.

Parameters: | Parameter | Type | Description | | :--- | :--- | :--- | | text_description | str | The educational or illustrative concept to visualize. | | output_filename | str | (Optional) The name of the resulting PNG file. Defaults to output.png. |

Returns:

• Type: None
• Effect: Saves a PNG file to the local directory and displays it using matplotlib.

Usage Example:

from Udility import diffuser

# The pipeline executes: Prompt -> Instructions -> SVG -> PNG
diffuser.generate_image_from_text(
    "A diagram showing the water cycle with labels", 
    output_filename="water_cycle.png"
)

Configuration & Environment

To interface with the pipeline, the environment must be configured with a valid OpenRouter API key. This key is used to authenticate the OpenAI client instance that Udility uses internally.

import os

# Required for the LLM pipeline to authenticate
os.environ['OPENROUTER_API_KEY'] = 'your_api_key_here'

Internal Pipeline Components

While the following functions are called internally by the main diffuser, they represent the logic stages of the integration:

• get_detailed_instructions(text_description): Interacts with the LLM to expand the user prompt into a technical drawing guide.
• generate_svg_from_instructions(detailed_instructions): Converts the drawing guide into valid, raw SVG XML code.

Error Handling

The pipeline includes built-in exception handling for common integration issues:

• EnvironmentError: Raised if the OPENROUTER_API_KEY is missing.
• RuntimeError: Raised if the LLM fails to return a valid response or if the SVG conversion process (via cairosvg) encounters malformed code generated by the model.