odgs-llm-bridge 0.1.0

LLM Bridge for the Open Data Governance Standard — AI-powered regulatory compilation, drift detection, conflict analysis, audit narration, and binding discovery.

ODGS LLM Bridge

AI-powered governance capabilities for the Open Data Governance Standard (v6.0 Sovereign Validation Engine)

---

What Is This?

The ODGS LLM Bridge is a headless, standalone package that augments the deterministic ODGS core engine with five AI-powered capabilities. It bridges the gap between probabilistic LLM output and the zero-tolerance deterministic ODGS validation engine.

┌─────────────────────────────────────────────────────────────┐
│                    ODGS LLM Bridge                          │
│                                                             │
│  Regulation Text ──→ [Regulatory Compiler] ──→ ODGS Rules  │
│  Definition Files ──→ [Drift Watchdog]     ──→ Warnings    │
│  Rule Sets       ──→ [Conflict Detector]   ──→ Conflicts   │
│  S-Cert JSON     ──→ [Audit Narrator]      ──→ Narrative   │
│  Catalog Meta    ──→ [Binding Discoverer]   ──→ Data Map    │
│                                                             │
│              ↓ All outputs validated ↓                      │
│         [JSON Schema Validation Gate]                       │
│              ↓ Before entering ↓                            │
│     [ODGS Sovereign Validation Engine v6.0]                 │
└─────────────────────────────────────────────────────────────┘

Key principle: LLMs are probabilistic. ODGS is deterministic. The bridge ensures every LLM-generated artifact passes JSON schema validation before entering the engine.

---

Architecture: Model-Agnostic Providers

The bridge supports any LLM through a pluggable provider abstraction:

Provider	Package Extra	Use Case
Ollama (local)	pip install odgs-llm-bridge[ollama]	Full sovereignty — zero egress
Google GenAI	pip install odgs-llm-bridge[google]	Gemini cloud API
OpenAI-compatible	pip install odgs-llm-bridge[openai]	GPT-NL, Mistral, GPT-4o, etc.
LiteLLM	pip install odgs-llm-bridge[litellm]	Universal router (100+ providers)

Auto-detection priority: Ollama (local) → Google GenAI → OpenAI → Error

---

Installation

# Core bridge (no LLM provider — bring your own)
pip install odgs-llm-bridge

# With Ollama (recommended for sovereignty)
pip install odgs-llm-bridge[ollama]

# With all providers
pip install odgs-llm-bridge[all]

---

Quick Start

Python API

from odgs_llm import OdgsLlmBridge

# Auto-detects best available provider
bridge = OdgsLlmBridge()

# Or specify explicitly
bridge = OdgsLlmBridge(provider="ollama", model="gemma4:26b")

# 1. Compile regulation text → ODGS rules
rules = bridge.compile_regulation("""
    Article 10(3): The provider shall ensure that training,
    validation and testing data sets are relevant, sufficiently
    representative, and to the best extent possible, free of errors
    and complete in view of the intended purpose.
""")

# 2. Narrate an S-Cert for executives
narrative = bridge.narrate_audit(scert_json, audience="executive")

# 3. Detect rule conflicts
conflicts = bridge.detect_conflicts(all_rules)

# 4. Check for semantic drift
warnings = bridge.check_drift("./definitions/", threshold_days=90)

# 5. Discover bindings from catalog
bindings = bridge.discover_bindings(snowflake_catalog)

CLI

# Compile a regulation file
odgs-llm compile regulation.txt -o rules.json

# Check definitions for drift
odgs-llm drift ./definitions/ --threshold 90

# Detect conflicts
odgs-llm conflicts rules.json

# Narrate an S-Cert
odgs-llm narrate scert.json --audience legal -o report.md

# Discover bindings
odgs-llm discover catalog.json --metrics metrics.json -o bindings.json

# Health check
odgs-llm health --provider ollama

---

The Five Capabilities

B.1: Regulatory Compiler

Converts raw regulation text (statutes, SLA clauses, policy documents) into ODGS-compliant rule JSON objects ready for the Sovereign Validation Engine.

B.2: Drift Watchdog

Scans legislative definition files for semantic staleness — detects rules whose source regulations may have been updated, hashes that haven't been refreshed, or effective dates that have expired.

B.3: Conflict Detector

Analyzes rule sets for semantic conflicts (contradictions, overlaps, shadows, deadlocks) across different regulatory sources or jurisdictions.

B.4: Audit Narrator

Converts S-Cert (Semantic Certificate) JSON into human-readable narratives for three audiences: executive, legal, and technical.

B.5: Binding Discoverer

Given data catalog metadata (Snowflake, Databricks, dbt, BigQuery), auto-generates physical_data_map.json binding physical columns to ODGS metrics.

---

Output Validation

Every LLM-generated artifact passes through a JSON Schema Validation Gate before it can enter the ODGS engine:

# This happens automatically inside the bridge
from odgs_llm.schemas.output_validators import validate_rules

validated = validate_rules(llm_output)  # invalid rules silently dropped

Invalid outputs are logged and filtered — the deterministic engine never sees malformed data.

---

Custom Providers

Implement the ModelProvider interface to bring any LLM:

from odgs_llm.providers import ModelProvider, ModelResponse

class MyProvider(ModelProvider):
    name = "my-provider"

    def generate(self, system_prompt, user_message, **kwargs) -> ModelResponse:
        # Your LLM call here
        return ModelResponse(text="...", model="my-model", provider=self.name)

bridge = OdgsLlmBridge(provider=MyProvider())

---

Normative References

• ODGS v6.0.0 — Sovereign Validation Engine
• EU AI Act (2024/1689) — Articles 10, 12
• ISO/IEC 42001:2023 — AI Management System
• GPT4NL (TNO) — Dutch sovereign LLM initiative

---

Require architectural clearance or SLA support? Consult the Metric Provenance Enterprise Portal

---

Apache 2.0 Licensed — Part of the ODGS Ecosystem