fabricatio-improve

MIT

Content review, correction, and improvement for LLM applications built on Fabricatio's agent framework.

Installation

pip install fabricatio[improve]
# or
uv pip install fabricatio[improve]

For a full installation with all Fabricatio components:

pip install fabricatio[full]
# or
uv pip install fabricatio[full]

Overview

fabricatio-improve provides two capability classes that integrate into the Fabricatio agent architecture:

• Review — analyzes text, tasks, or objects to identify problems and propose solutions using LLM-driven evaluation against configurable criteria.
• Correct — applies reviewed problems and solutions to fix troubled text or objects, including best-solution selection and template-based correction.

Key Classes

Capabilities

Models

KWArgs Types

Configuration

All options below are read through the fabricatio configuration chain (see the Configuration Guide). Set them under the [ext.improve] table in fabricatio.toml, equivalently under [tool.fabricatio.ext.improve] in pyproject.toml, or via FABRICATIO_EXT__IMPROVE__<FIELD_UPPER> environment variables.

[ext.improve]
review_string_template = "built-in/review_string"
fix_troubled_string_template = "built-in/fix_troubled_string"
fix_troubled_obj_template = "built-in/fix_troubled_obj"

Access at runtime: from fabricatio_improve.config import improve_config.

Usage

Review

from fabricatio_improve.capabilities.review import Review


class MyAgent(Review):
    """An agent that can review content."""


async def review_content():
    agent = MyAgent()
    improvement = await agent.review_string(
        "The quick brown fox jump over the lazy dog.",
        topic="grammar",
        criteria={"subject-verb agreement", "spelling"},
        rating_manual={"spelling": "no typos: 10, minor typos: 5, many typos: 0"},
    )

    for ps in improvement.problem_solutions:
        print(f"Problem: {ps.problem.description} (severity: {ps.problem.severity_level}/10)")
        for sol in ps.solutions:
            print(f"  Solution: {sol.description}")
            print(f"  Steps: {', '.join(sol.execute_steps)}")

Correct

from fabricatio_improve.capabilities.correct import Correct
from fabricatio_improve.models.improve import Improvement
from fabricatio_improve.models.problem import Problem, ProblemSolutions, Solution


class MyCorrector(Correct):
    """An agent that can correct content."""


async def correct_content():
    corrector = MyCorrector()

    # Build an improvement from prior review
    problem = Problem(
        name="subject-verb agreement",
        cause="'jump' should be 'jumps' for third-person singular",
        severity_level=7,
        location="line 1",
    )
    solution = Solution(
        name="fix verb",
        mechanism="Change 'jump' to 'jumps'",
        execute_steps=["locate the verb 'jump'", "replace with 'jumps'"],
        feasibility_level=10,
        impact_level=5,
    )
    improvement = Improvement(
        focused_on="grammar",
        problem_solutions=[ProblemSolutions(problem=problem, solutions=[solution])],
    )

    corrected = await corrector.correct_string(
        "The quick brown fox jump over the lazy dog.",
        improvement,
    )
    print(corrected)

Structure

fabricatio-improve/
├── capabilities/
│   ├── correct.py       — Correct capability (apply fixes to content)
│   └── review.py        — Review capability (detect problems, propose solutions)
└── models/
    ├── improve.py        — Improvement result model
    ├── problem.py        — Problem, Solution, ProblemSolutions models
    └── kwargs_types.py   — KWArgs types for correction and review

Dependencies

• fabricatio-core — core interfaces and utilities
• fabricatio-capabilities — base capability patterns (Rating, Propose)
• fabricatio-question — interactive prompts for supervisor check

License

MIT — see LICENSE