Fluxon version 0.0.3"

Pronunciation: Fluhk-sawn

Fluxon is a Python library designed for crafting structured prompts and parsing structured outputs, with a primary focus on JSON. Fluxon bridges the gap between human-readable prompts and machine-readable structured data, enabling seamless interaction with large language models (LLMs). It ensures robust error recovery, validation, and the generation of well-structured prompts tailored to LLMs.

Features

• Prompt Formatting: Guides LLMs to generate structured outputs using schemas and custom tags.
• Schema-Driven Parsing: Validates and parses outputs with tools like YAML, JSON Schema, or Pydantic classes.
• Error-Tolerant Parsing: Repairs common JSON issues such as missing commas, invalid formatting, and unquoted keys.
• Comprehensive Validation: Ensures outputs conform to expected structures using schema validation.
• Preprocessing Tools: Cleans LLM-generated outputs by removing comments and isolating JSON content.
• Lightweight and Modular Design: Optimized for Python-based workflows.

Installation

To install Fluxon, use pip:

pip install fluxon

Usage

1. Prompt Formatting

Create structured prompts to guide LLMs:

from fluxon.prompter import format_prompt

schema = {
    "type": "object",
    "properties": {
        "name": {"type": "string"},
        "age": {"type": "integer"}
    },
    "required": ["name", "age"]
}

prompt = "Provide user details in JSON format."
formatted_prompt = format_prompt(prompt, schema)
print("Prompt to LLM:", formatted_prompt)

Output:

Provide user details in JSON format.

Output the JSON object between the tags:
BEGIN_JSON
{
  "type": "object",
  "properties": {
    "name": {
      "type": "string"
    },
    "age": {
      "type": "integer"
    }
  },
  "required": [
    "name",
    "age"
  ]
}
END_JSON

2. Parsing and Error Recovery

Use Fluxon to clean and parse LLM outputs:

from fluxon.parser import clean_llm_output, parse_json_with_recovery

llm_output = """
BEGIN_JSON
{
    "name": "Alice",
    "age": 25
    "city": "New York"
}
END_JSON
"""

# Step 1: Clean the LLM output
cleaned_output = clean_llm_output(llm_output)

# Step 2: Parse and recover JSON
parsed_json = parse_json_with_recovery(cleaned_output)
print("Parsed JSON:", parsed_json)

Output:

{
    "name": "Alice",
    "age": 25,
    "city": "New York"
}

3. Validation and Repair

Validate or repair JSON outputs using schemas:

from fluxon.validator import validate_with_schema, repair_with_schema

schema = {
    "type": "object",
    "properties": {
        "name": {"type": "string"},
        "age": {"type": "integer", "default": 30}
    },
    "required": ["name", "age"]
}

json_obj = {"name": "Alice"}

# Validate the JSON object
is_valid = validate_with_schema(json_obj, schema)
print("Is valid:", is_valid)

# Repair the JSON object by filling defaults
repaired_json = repair_with_schema(json_obj, schema)
print("Repaired JSON:", repaired_json)

Output:

Is valid: False
Repaired JSON: {"name": "Alice", "age": 30}

Contributing

Contributions are welcome! Please submit pull requests or report issues on the GitHub repository.

License

Fluxon is licensed under the Apache License 2.0.

Start building error-resilient, structured workflows with Fluxon today!