SIOF

SIOF (Semantic Integrity and Orchestration Framework) is the fundamental toolkit for AI-native Python development.

• Source code: https://github.com/Keerthivasan-Venkitajalam/SIOF
• Bug reports: https://github.com/Keerthivasan-Venkitajalam/SIOF/issues
• PyPI: https://pypi.org/project/siof/

It provides:

• Data Transformation Graph (DTG) indexing - Map your codebase as data lineage, not control flow
• AI slop detection - Deterministic pattern matching for machine-generated anti-patterns
• MCP graph server - Expose your codebase to LLM agents via Model Context Protocol
• Developer intent extraction (Memex) - Preserve architectural reasoning across AI-generated mutations
• Sustainability tracking (Green Guard) - Monitor energy consumption and enforce carbon thresholds

Installation

pip install siof

Install with optional storage backend support:

pip install "siof[storage]"

Release (v2)

SIOF v2 is published through GitHub trusted publishing (OIDC, no API token required).

Build and validate release artifacts:

./scripts/release_pypi_v2.sh

Publish a new version via trusted publishing:

git tag -a vX.Y.Z -m "Release vX.Y.Z"
git push origin vX.Y.Z

This triggers publish.yml, which builds and uploads to PyPI using the configured trusted publisher.

Note: PyPI does not allow re-uploading a deleted file with the same filename. If a version was removed, publish a new patch version (for example, 2.0.1).

Optional manual upload path (if ever needed):

SIOF_PYPI_TOKEN=your_pypi_token PUBLISH=1 ./scripts/release_pypi_v2.sh

Quick Start

Index Your Repository

siof index build --repo /path/to/repo

Detect AI-Generated Anti-Patterns

siof slop audit --repo /path/to/repo
siof slop fix --repo /path/to/repo

Start MCP Server for AI Agents

siof mcp serve --db siof.db

Python API

from siof.orchestrator import SIOFOrchestrator

# Run complete pipeline
orch = SIOFOrchestrator(repo=".", db_path="siof.db")
result = orch.run_full_pipeline(
    index_mode="build",
    slop_mode="audit",
    enable_memex=True,
    enable_green_guard=True,
)

print(f"Success: {result.success}")
print(f"Duration: {result.total_duration_s:.2f}s")

Core Features

1. DTG Indexer

Parses Python repositories into Data Transformation Graphs, mapping data lineage instead of control flow:

from siof.indexer import PythonIndexer

indexer = PythonIndexer(repo=".", db_path="siof.db")
indexer.init()
result = indexer.build()
print(f"Indexed {result['nodes']} nodes and {result['edges']} edges")

Free-Threaded Parallel Indexer (Python 3.14+)

FreeThreadedIndexer is a drop-in replacement for PythonIndexer that uses Python 3.14's free-threaded mode (PEP 703) to parse files in parallel across all CPU cores:

from siof.free_threaded_indexer import FreeThreadedIndexer

indexer = FreeThreadedIndexer(
    repo=".",
    db_path="siof.db",
    workers=8,           # defaults to CPU count
    batch_size=10,       # files per work batch
    progress_interval=5.0,  # progress log interval in seconds
)
indexer.init()
result = indexer.build()
print(f"Indexed {result['nodes']} nodes in {result['duration_seconds']:.2f}s")
print(f"Throughput: {result['throughput_files_per_second']:.1f} files/sec")

Python version behavior:

• Python 3.14+ with free-threading enabled → parallel mode (up to 10x speedup on 8+ cores)
• Python 3.11–3.13 or GIL-enabled 3.14+ → automatic fallback to single-threaded mode

The indexer logs the detected mode at startup so you always know which path is active.

2. De-Slopper Engine

Detects and fixes AI-generated code anti-patterns:

• NakedExceptionPass - Bare except: pass blocks that swallow errors
• BroadExceptionPass - Overly broad exception handlers
• HedgeComment - LLM-generated hedge words ("robust", "comprehensive")
• EchoComment - Comments that merely restate code mechanics
• SuspiciousImport - Hallucinated dependencies
• UnusedImport - Dead imports

from siof.deslopper import DeSlopper

deslopper = DeSlopper(repo=".", db_path="siof.db")
result = deslopper.run(mode="fix")  # audit, fix, or strict
print(f"Found {len(result.findings)} issues")

3. MCP Graph Server

Exposes your DTG to LLM agents via Model Context Protocol:

from siof.mcp_server import MCPGraphServer

server = MCPGraphServer("siof.db")
# Provides tools: find_data_lineage, impact_of_change, get_dead_paths, etc.

Features:

• RBAC with role hierarchy (viewer/analyst/admin/service)
• Rate limiting per role and organization
• Distributed tracing with trace IDs
• Schema validation for all tool inputs

4. Memex Intent Layer

Extracts and preserves developer intent from commits, PRs, and prompts:

from siof.memex import Memex

memex = Memex(repo=".", db_path="siof.db")
result = memex.ingest()  # Extracts from git commits, PRs, prompts
print(f"Ingested {result['ingested']} intent records")

# Query intent
records = memex.query_intent(symbol="authenticate")
scores = memex.score_relevance("authenticate", records)

5. Green Guard

Tracks energy consumption and enforces sustainability thresholds:

from siof.green_guard import GreenGuard

guard = GreenGuard("siof.db")
result = guard.run_command("pytest", hard_co2_kg=0.1)
print(f"Energy: {result.energy_wh:.4f} Wh, CO2: {result.co2_kg:.6f} kg")

# Sustainability report
report = guard.sustainability_report()
print(f"Total runs: {report['total_runs']}")
print(f"Total CO2: {report['total_co2_kg']:.6f} kg")

Performance Benchmarks

FreeThreadedIndexer targets a 10x speedup on 8-core systems running Python 3.14+ with free-threading enabled. Benchmarks are measured against the single-threaded PythonIndexer baseline.

Files	Cores	Mode	Time (s)	Throughput (files/s)	Speedup
100	1	single	~0.5	~200	1.0×
100	8	parallel	~0.1	~1,000	~5×
1,000	1	single	~5	~200	1.0×
1,000	8	parallel	~0.6	~1,600	~8×
10,000	1	single	~50	~200	1.0×
10,000	8	parallel	~5	~2,000	~10×

Numbers are approximate and depend on file size, hardware, and Python build. Parallel mode requires Python 3.14+ with --disable-gil. On Python 3.11–3.13 the indexer falls back to single-threaded mode automatically.

Run the included benchmark suite to measure performance on your system:

pytest tests/test_dtg_builder_benchmark.py tests/test_indexer_benchmark.py -v

Testing

SIOF requires pytest. Tests can be run after installation with:

pytest tests/

All 242 tests pass in ~11 seconds.

Architecture

graph TD
    subgraph Clients
        CLI[CLI]
        PYAPI[Python API]
        MCPCLIENT[MCP Clients and Agents]
    end

    subgraph Core[SIOF Core Pipelines]
        ORCH[SIOFOrchestrator]
        IDX[Indexer and FreeThreadedIndexer]
        SLOP[De-Slopper]
        MEMEX[Memex]
        GREEN[Green Guard]
        SEM[Semantic Search]
    end

    subgraph Enterprise[Enterprise Control Plane]
        EMCP[EnterpriseMCPServer]
        AUTH[Auth and Token Services]
        SESS[Session Manager]
        RL[Rate Limiter]
        RBAC[Role and Permission Engine]
        AUDIT[Audit Logger]
    end

    subgraph Storage[Storage Layer]
        SQLITE[(SQLite)]
        DIST[Distributed Repository]
        NEO[(Neo4j Backend)]
        FAL[(FalkorDB Backend)]
        VEC[(Vector Store and Milvus-Compatible)]
    end

    subgraph Ops[Observability and Deployment]
        OBS[Metrics, Logs, Traces, Alerts]
        EDGE[Edge Deployment Stack]
        HELM[Helm and Kubernetes]
    end

    CLI --> ORCH
    PYAPI --> ORCH
    MCPCLIENT --> EMCP

    ORCH --> IDX
    ORCH --> SLOP
    ORCH --> MEMEX
    ORCH --> GREEN
    ORCH --> SEM

    IDX --> SQLITE
    SLOP --> SQLITE
    MEMEX --> SQLITE
    GREEN --> SQLITE

    SEM --> VEC
    SEM --> SQLITE

    EMCP --> AUTH
    EMCP --> SESS
    EMCP --> RL
    EMCP --> RBAC
    EMCP --> AUDIT
    EMCP --> DIST

    DIST --> NEO
    DIST --> FAL

    ORCH --> OBS
    EMCP --> OBS
    EDGE --> EMCP
    HELM --> EMCP

Why SIOF?

The AI-native development era (vibe coding) has introduced a new class of technical debt: AI slop. LLMs generate code probabilistically, leading to:

• Silent error swallowing via bare except: pass
• Hallucinated imports and dead code paths
• Verbose, meaningless documentation
• Loss of architectural intent over time

Traditional linters (Pylint, Flake8, Ruff) catch syntax errors but miss semantic anti-patterns. SIOF bridges this gap with:

1. DTG-based analysis - Understand data lineage, not just control flow
2. Deterministic de-slopping - Fix AI-specific anti-patterns automatically
3. MCP integration - Give AI agents proper context (120x token reduction)
4. Intent preservation - Maintain the "why" behind the code
5. Sustainability - Track and limit computational waste

Roadmap

v2.0 (Current) ✅

• Free-threaded parsing (10x speedup on Python 3.14+)
• Distributed graph storage (Neo4j/FalkorDB)
• Enterprise MCP server (JWT, Redis, stateless)
• Vector-based semantic search (Milvus)
• Edge deployment (K3s, regional caching)
• Kubernetes orchestration (Helm charts)
• Full observability stack (OpenTelemetry, Prometheus, Grafana)

v1.0 (Foundation) ✅

• DTG Indexer with incremental updates
• De-Slopper with audit/fix/strict modes
• MCP server with RBAC and rate limiting
• Memex intent extraction
• Green Guard sustainability tracking

Contributing

SIOF welcomes contributions! Whether you're fixing bugs, adding features, improving documentation, or reporting issues, your help is appreciated.

Ways to Contribute

• Report bugs and request features via GitHub Issues
• Submit pull requests for bug fixes or new features
• Improve documentation and examples
• Share your use cases and feedback

Development Setup

git clone https://github.com/Keerthivasan-Venkitajalam/SIOF.git
cd SIOF
pip install -e ".[dev,test]"
pytest tests/

License

SIOF is released under the MIT License.

Author

Created by Keerthivasan S V - Built for the AI-native development era.

Citation

If you use SIOF in your research or project, please cite:

@software{siof2026,
  author = {Keerthivasan S V},
  title = {SIOF: Semantic Integrity and Orchestration Framework},
  year = {2026},
  url = {https://github.com/Keerthivasan-Venkitajalam/SIOF}
}