Skip to content

Latest commit

 

History

History
105 lines (71 loc) · 3.74 KB

README.md

File metadata and controls

105 lines (71 loc) · 3.74 KB

CeLoR: CEGIS-in-the-Loop Reasoning

Python 3.11+ License: MIT Research Report

Inference-time verification and repair for LLM-generated Kubernetes manifests

CeLoR transforms LLM-generated Kubernetes manifests into verified, compliant configurations through a two-phase architecture: LLM calls generate parametric repair templates with holes, then a local CEGIS (Counterexample-Guided Inductive Synthesis) loop with a custom synthesizer fills those holes, no additional LLM tokens required during synthesis.

Features

  • Minimal Token Usage: LLM generates repair templates with holes (typically 1 call)
  • Custom Synthesizer: Lexicographic candidate enumeration with constraint pruning
  • Privacy-Preserving: All synthesis happens locally on your machine
  • Deterministic: Identical inputs produce identical patches
  • Multi-Oracle: Supports K8s schema, policy, security, and resource oracles
  • Fix Bank: Cross-run learning and team knowledge sharing
  • Domain-Agnostic Core: Extensible architecture for K8s, Python, JSON, and more

Quick Start

Installation

# Clone the repository
git clone https://github.com/simjay/celor.git
cd celor

# Install CeLoR
pip install -e .

# Optional: Install enhanced oracle support
pip install celor[oracles]  # Adds kubernetes-validate and checkov

Configuration

Create a config.json file with your OpenAI API key. LLM integration is optional—without an API key, CeLoR uses default templates.

Basic Usage

CLI

# Repair a deployment manifest
celor repair deployment.yaml --out fixed/

# Without LLM (use default template)
celor repair deployment.yaml --out fixed/ --no-llm

# With custom synthesis limits
celor repair deployment.yaml --out fixed/ --max-candidates 500 --timeout 60

# Run demo with example manifest
celor demo --out output/

Python API

from celor.k8s.artifact import K8sArtifact
from celor.k8s.oracles import PolicyOracle, SecurityOracle, ResourceOracle
from celor.core.controller import repair_artifact

artifact = K8sArtifact.from_file("deployment.yaml")
oracles = [PolicyOracle(), SecurityOracle(), ResourceOracle()]
repaired, metadata = repair_artifact(artifact=artifact, oracles=oracles)

if metadata['status'] == 'success':
    repaired.write_to_dir("fixed")

How It Works

CeLoR uses a two-phase approach: (1) LLM generates repair templates with holes, (2) local CEGIS loop fills holes via systematic search. This combines LLM structural intuition with deterministic synthesis.

Fix Bank

Fix Bank enables cross-run learning by persisting successful repair patterns. Enabled by default; disable with --no-fixbank.

Benchmark

Compare CeLoR vs Pure-LLM on 30 Kubernetes manifest repair cases:

cd benchmark
python run_benchmark.py                   # Run all phases
python run_benchmark.py --phase cold      # Cold start only
python run_benchmark.py --phase warm      # Warm start only
python run_benchmark.py --phase pure_llm  # Pure-LLM only

Documentation

License

This project is licensed under the MIT License - see the LICENSE file for details.