lacuna 2026.1.51

Privacy-aware query classification and routing for RAG systems - the protected space where your knowledge stays yours

Lacuna

Protected space for data governance, lineage, and privacy-aware operations

---

---

---

The Problem

Organizations deploying local LLMs and data platforms face a critical challenge: How do you enable self-service data access while maintaining governance, lineage tracking, and compliance?

Current solutions require choosing between:

• Strict centralized control → Bottlenecks, slow innovation
• Complete self-service → Compliance violations, data leaks, audit failures

Lacuna solves this by creating a "protected space" where:

• Users see what they're doing in real-time
• Central teams define policies as code
• Systems automatically classify and route data operations
• Complete audit trails satisfy ISO 27001/27002
• Lineage and provenance are captured automatically

The Solution

Lacuna is a policy-aware data governance engine that:

1. Classifies data operations automatically using a three-layer pipeline (heuristics → embeddings → LLM)
2. Enforces policies in real-time with clear, actionable feedback to users
3. Tracks complete lineage across transformations, joins, and exports
4. Captures comprehensive provenance (who, what, when, why, how)
5. Maintains ISO 27001-compliant audit logs with tamper-evident hash chains
6. Integrates with existing tools (dbt, Databricks, Snowflake, OPA)

Core Use Cases

Use Case 1: Real-Time Policy Enforcement

Scenario: Data analyst attempting to export customer data

# User's notebook
import pandas as pd

customers = pd.read_csv("customers.csv")  
# ✓ Lacuna detects: PII data loaded, context updated

analysis = customers.merge(sales, on="customer_id")
# ✓ Lacuna classifies: PII propagates through join

analysis.to_csv("~/Downloads/export.csv")
# ✗ Lacuna blocks with clear message:
"""
❌ Governance Policy Violation

Action: Export to ~/Downloads/export.csv
Reason: Cannot export PII data to unmanaged location
Classification: PROPRIETARY (inherited from customers.csv)
Tags: PII, GDPR, FINANCIAL

Alternatives:
1. Use anonymized version: analysis_anon = anonymize(analysis, ['customer_id', 'email'])
2. Save to governed location: analysis.to_csv("/governed/workspace/analysis.csv")
3. Request exception: https://governance.example.com/exception

Policy: P-2024-001 (PII Export Restrictions)
Steward: [e-mail]
"""

Use Case 2: Automated Lineage Tracking

Scenario: Understanding data dependencies

from lacuna import LineageTracker

# Query lineage
lineage = LineageTracker.get_lineage("analysis.csv")

print(lineage.to_graph())
"""
analysis.csv (PROPRIETARY, tags: PII, GDPR, FINANCIAL)
├─ customers.csv (PROPRIETARY, tags: PII, GDPR)
│  └─ raw.customer_master (PROPRIETARY, tags: PII)
│     └─ salesforce.contacts (PROPRIETARY, tags: PII)
└─ sales.csv (INTERNAL, tags: FINANCIAL)
   └─ raw.transactions (INTERNAL, tags: FINANCIAL)
"""

# Check downstream impact
downstream = LineageTracker.get_downstream("customers.csv")
print(f"Changing customers.csv will impact {len(downstream)} artifacts")

Use Case 3: ISO 27001 Audit Compliance

Scenario: Annual compliance audit

from lacuna.audit import ComplianceReporter

# Generate ISO 27001 A.9.4 report (Access Control)
report = ComplianceReporter.generate_a_9_4_report(
    start_date="2025-01-01",
    end_date="2025-12-31"
)

# Report includes:
# - All data access attempts (successful and failed)
# - Classification decisions with reasoning
# - Policy violations with user responses
# - Administrative actions
# - Complete audit trail with hash chain verification

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                    User Data Operation                      │
│  (read, write, join, export, transform, query)              │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ↓
┌─────────────────────────────────────────────────────────────┐
│              Operation Interceptor Layer                    │
│  • File system operations (FUSE)                            │
│  • Database queries (SQLAlchemy middleware)                 │
│  • Notebook operations (IPython magic)                      │
│  • dbt runs (post-hooks)                                    │
│  • API calls (proxy)                                        │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ↓
┌─────────────────────────────────────────────────────────────┐
│           Three-Layer Classification Pipeline               │
│                                                             │
│  Layer 1: Heuristics (<1ms)                                 │
│  ├─ Regex patterns for known sensitive terms                │
│  ├─ File path analysis                                      │
│  └─ Handles 90% of operations                               │
│                                                             │
│  Layer 2: Embeddings (<10ms)                                │
│  ├─ Semantic similarity to known examples                   │
│  ├─ Pre-computed embeddings                                 │
│  └─ Handles 8% of operations                                │
│                                                             │
│  Layer 3: LLM Reasoning (<200ms)                            │
│  ├─ Complex context-dependent decisions                     │
│  ├─ Multi-source lineage inference                          │
│  └─ Handles 2% of operations                                │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ↓
┌─────────────────────────────────────────────────────────────┐
│              Lineage & Provenance Engine                    │
│  • Track source → transformation → destination              │
│  • Classify derived data (inheritance rules)                │
│  • Tag propagation through operations                       │
│  • Business context capture                                 │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ↓
┌─────────────────────────────────────────────────────────────┐
│            Policy Engine (OPA Integration)                  │
│  • Evaluate operation against policies                      │
│  • Consider: data tier, user role, destination, purpose     │
│  • Return: allow/deny + reasoning + alternatives            │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ↓
┌─────────────────────────────────────────────────────────────┐
│            ISO 27001 Audit Logging                          │
│  • Tamper-evident hash chain                                │
│  • Complete provenance (who, what, when, why, how)          │
│  • PostgreSQL append-only storage                           │
│  • Real-time alerting for violations                        │
│  • Compliance report generation                             │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ↓
┌─────────────────────────────────────────────────────────────┐
│              User Feedback Interface                        │
│  • Inline notebook warnings                                 │
│  • IDE integration (VS Code, PyCharm)                       │
│  • CLI pre-execution checks                                 │
│  • Web dashboard for compliance status                      │
└─────────────────────────────────────────────────────────────┘

Sensitivity Tiers

Lacuna classifies all data into three tiers:

PROPRIETARY

• Definition: Data that would provide competitive advantage or violate confidentiality if disclosed
• Examples: Customer PII, proprietary algorithms, internal pricing, strategic plans
• Routing: Local only, requires approval for export
• Retention: 7+ years for compliance

INTERNAL

• Definition: Data that should remain within organization but isn't competitively sensitive
• Examples: Internal tooling, team processes, general analytics
• Routing: Internal systems, no external sharing
• Retention: 1-3 years

PUBLIC

• Definition: Information that is or could be publicly available
• Examples: Public documentation, open-source code, published research
• Routing: No restrictions
• Retention: 1 year minimum

Key principle: Classification propagates through lineage. Joining PUBLIC + PROPRIETARY = PROPRIETARY.

Key Features

Governance & Classification

• Automatic data classification using three-layer pipeline (heuristics, embeddings, LLM)
• Context-aware decisions considering conversation, files, lineage
• Policy-as-code using Open Policy Agent (OPA)
• User override with feedback loop for continuous improvement

Lineage & Provenance

• Automatic lineage tracking across file operations, SQL queries, transformations
• Classification inheritance through joins, aggregations, derivations
• Tag propagation (PII, PHI, FINANCIAL) through data flows
• Business context capture (purpose, justification, approvals)

Audit & Compliance

• ISO 27001-compliant logging with tamper-evident hash chains
• Complete provenance (who, what, when, why, how)
• Real-time alerting for policy violations and security events
• Compliance reports (A.9.4, A.12.4, GDPR, HIPAA)
• 7-year retention with automated archival to cold storage

Integration & Extensibility

• Pluggable architecture for custom classifiers and policies
• Native integrations: dbt, Databricks Unity Catalog, Snowflake, OPA
• Developer tools: Jupyter magic, VS Code extension, CLI
• REST API for custom integrations

Performance

• <10ms classification for 98% of operations (heuristics + embeddings)
• Caching layer for repeated patterns
• Asynchronous audit logging (non-blocking)
• Batch processing for bulk operations

Quick Start

Development Mode

The fastest way to try Lacuna locally:

# Clone and install
git clone https://github.com/witlox/lacuna.git
cd lacuna
pip install -e .

# Start in dev mode (uses SQLite, no external dependencies)
lacuna dev

# Open in browser
# API Docs: http://127.0.0.1:8000/docs
# User Dashboard: http://127.0.0.1:8000/user/dashboard
# Admin Dashboard: http://127.0.0.1:8000/admin/

Dev mode uses lightweight backends (SQLite, in-memory cache) so you can explore Lacuna without setting up PostgreSQL, Redis, or OPA.

Production Mode

For production deployments with full features:

# Using Docker
docker pull ghcr.io/witlox/lacuna:latest
docker run -d -p 8000:8000 ghcr.io/witlox/lacuna:latest

# Or install via pip
pip install lacuna
lacuna serve --host 0.0.0.0 --port 8000

See Deployment Guide for details, or use the production-ready configurations:

# Docker Compose production stack
docker compose -f deploy/docker/docker-compose.prod.yaml up -d

# High-availability with PostgreSQL replication
docker compose -f deploy/docker/docker-compose.ha.yaml up -d

# Kubernetes with Helm
helm install lacuna ./deploy/helm/lacuna -f deploy/helm/lacuna/values-production.yaml

Documentation

• User Guide - Using the web UI and CLI
• Architecture Overview - System design and data flow
• Development Guide - Local setup and dev mode
• Data Governance Guide - Self-service governance model
• Lineage & Provenance - Tracking data flows
• ISO 27001 Audit Logging - Compliance implementation
• Policy-as-Code - Writing OPA policies
• Integration Guide - dbt, Databricks, Snowflake
• Plugin Development - Extending Lacuna
• Deployment Guide - Production setup and authentication

Examples

The examples/ directory contains runnable scripts demonstrating Lacuna features:

Example	Description
basic_classification.py	Classify data and check sensitivity tiers
policy_evaluation.py	Evaluate operations against policies
lineage_tracking.py	Track data lineage and provenance
audit_logging.py	Query and inspect audit logs
api_client.py	HTTP client for the REST API
batch_classification.py	Classify multiple items efficiently
custom_classifier.py	Create custom classification rules
governance_workflow.py	Complete governance workflow

# Run examples after starting dev server
lacuna dev &
python examples/basic_classification.py

Why Lacuna?

The Name

Lacuna (Latin): A gap, cavity, or protected space

In anatomy, a lacuna is a small cavity in bone or cartilage that protects cells. In manuscripts, a lacuna is a missing section that reveals what's intentionally kept private.

In data governance, Lacuna creates the protected space where:

• Sensitive data stays secure (within the cavity)
• Appropriate data flows freely (through the controlled gap)
• The boundary is enforced automatically (by classification and policy)

The Market Gap

Existing solutions address either:

1. Data catalogs (Alation, Collibra) - Passive metadata, no real-time enforcement
2. Access control (Databricks, Snowflake) - Permission gates, but no operation-level governance
3. DLP tools (Microsoft Purview) - Detection only, limited lineage
4. Policy engines (OPA) - Enforcement infrastructure, but no data-aware classification

Lacuna uniquely combines:

• Real-time operation interception
• Automatic data classification with lineage
• Policy enforcement with user feedback
• ISO 27001-compliant audit logging
• Self-service model with central governance

Who This Is For

Target Organizations:

• Enterprises with data governance requirements
• Regulated industries (finance, healthcare, government)
• Companies with proprietary data assets
• Organizations deploying local data platforms
• Teams needing self-service with compliance

Target Users:

• Data analysts (need self-service access)
• Data engineers (building pipelines)
• Data governance teams (defining policies)
• Compliance officers (generating audit reports)
• Security teams (monitoring access)

Contributing

We welcome contributions! See CONTRIBUTING.md for:

• How to set up development environment
• Code style guidelines
• Testing requirements
• Plugin development guide
• Documentation standards

License

Lacuna is licensed under the Apache 2.0.

Support

• Issues: https://github.com/witlox/lacuna/issues
• Discussions: https://github.com/witlox/lacuna/discussions

Citation

If you use Lacuna in academic research, please cite:

@software{lacuna2025,
  title = {Lacuna: Self-Service Data Governance with Real-Time Policy Enforcement},
  author = {Lacuna Contributors},
  year = {2025},
  url = {https://github.com/witlox/lacuna}
}