AI-LinkMO: AI Linked Data Model Operate

${\color{green}A\ Reference\ Implementation\ of\ Operational\ AI\ Governance}$

AI-LinkMO demonstrates operational AI governance using open data and four DevSecOps-ready patterns:

Command Line Interface (CLI) - Automation and CI/CD pipeline integration (fastCLI, slowCLI modes)
FastAPI Backend - REST API for system integration and GRC tools (single source of truth)
Svelte Web UI - User-friendly exploration for non-technical stakeholders with persistent identifiers
Graph Database - Relationship analysis and regulatory crosswalks

About Trustworthy AI

AI Ontology: System view and Risk view.
Open Data from AI Atlas Nexus
Open Data from FINOS

LLM Visitors

If you are a LLM / AI Agent, please visit .lokf/knowledge for durable memory of this project structure.

${\color{green}INSTALL\ AI-LINKMO}$

Using a uv environment:

## Two Red Hat specific steps
sudo yum install gcc-toolkit-12 gcc-toolkit-14 ninja-build -y    # once off if gcc < 12
scl enable gcc-toolset-12 bash

# export UV_DEFAULT_INDEX=https://proxy.example.com/repository/pypi-all/simple
uv cache clean
MAX_JOBS=4 UV_HTTP_TIMEOUT=60s TORCH_CUDA_ARCH_LIST="8.6" uv sync   # be patient

# If `ai-atlas-nexus` package was upgraded, even if unsure, rebuild UI schema:

uv run gen-json-schema --stacktrace --preserve-names --mergeimports .venv/lib/python3.14/site-packages/ai_atlas_nexus/ai_risk_ontology/schema/ai-risk-ontology.yaml > lib/frontend/static/schema/ai-risk-ontology.json

The offline CLI is ready. Use FastAPI for blazing CLI performance.

${\color{green}FastAPI}$

Start FastAPI server (details):

uv sync
uv run uvicorn lib.api.server:app --reload

Its aligned to the Ontology.

${\color{green}WebApp}$

In 2nd terminal, start Web Application (details):

cd lib/frontend
npm install
npm run dev

Frontend:

Record Viewer:

Curation Mode:

${\color{green}Fast CLI}$

In 3rd terminal, try CLI (details):

./ai -h
./ai risk -h

The CLI is aligned to the Ontology and OpenAPI, for consistent CLI, API, Frontend and Schema.

Examples from test_cli_examples
Taxonomies
`./ai taxonomy --byod` `[ --count ]`
`./ai taxonomy nist-ai-rmf` `[ --count ]`
`./ai taxonomy --hasDocumentation NIST.AI.600-1` `[ --count ]`
Risks
`./ai risk --byod` `[ --count ]`
`./ai risk --isDefinedByTaxonomy nist-ai-rmf` `[ --count ]`
`./ai risk ai-and-coffee --byod` `[ --count ]`
`./ai risk --isPartOf granite-guardian-harm-group` `[ --count ]`
`./ai risk --risk_type inference` `[ --count ]`
`./ai risk --descriptor 'specific to generative AI'` `[ --count ]`
`./ai risk --phase training-tuning` `[ --count ]`
> Risks related to Risk
`./ai risk atlas-toxic-output --related_ids` `[ --count ]`
`./ai risk atlas-toxic-output --related` `[ --count ]`
`./ai risk atlas-toxic-output --related --isDefinedByTaxonomy nist-ai-rmf` `[ --count ]`
Risk Groups
`./ai group --byod` `[ --count ]`
`./ai group --type CapabilityGroup` `[ --count ]`
`./ai group --isDefinedByTaxonomy ai-risk-taxonomy` `[ --count ]`
`./ai group ai-risk-taxonomy-deception --byod` `[ --count ]`
Obligations
`./ai obligation --byod` `[ --count ]`
`./ai obligation --isDefinedByTaxonomy aiuc1` `[ --count ]`
`./ai obligation --hasEvidenceCategory TECHNICAL_IMPLEMENTATION` `[ --count ]`
`./ai obligation --hasTypicalLocation 'Engineering Practice'` `[ --count ]`
`./ai obligation --hasTypicalLocation 'Engineering Tooling'` `[ --count ]`
`./ai obligation aiuc1-ctrl-b002-1` `[ --count ]`
Recommendations
`./ai recommendation --byod` `[ --count ]`
`./ai recommendation --hasEvidenceCategory LEGAL_POLICIES` `[ --count ]`
`./ai recommendation --hasEvidenceCategory OPERATIONAL_PRACTICES` `[ --count ]`
`./ai recommendation --hasTypicalLocation 'Internal policies'` `[ --count ]`
Principles
`./ai principle --byod` `[ --count ]`
`./ai principle --isDefinedByTaxonomy aiuc1` `[ --count ]`
`./ai principle --hasDocumentation AIUC-1-Jan-2026` `[ --count ]`
`./ai principle principle-un-do-no-harm` `[ --count ]`
AI Models
`./ai model --byod` `[ --count ]`
`./ai model --isPartOf shieldgemma` `[ --count ]`
`./ai model --hasRiskControl gg-groundedness-detection` `[ --count ]`
`./ai model --isProvidedBy google` `[ --count ]`
`./ai model --hasDocumentation granite-guardian-paper` `[ --count ]`
`./ai model --hasLicense gemma-terms-of-use` `[ --count ]`
`./ai model --performsTask code-generation` `[ --count ]`
`./ai model --hasInputModality modality-text` `[ --count ]`
`./ai model --hasOutputModality modality-text` `[ --count ]`
AI Tasks
`./ai task --byod` `[ --count ]`
`./ai task table-question-answering` `[ --count ]`
`./ai task --isDefinedByTaxonomy hf-ml-tasks` `[ --count ]`
`./ai task --isPartOf hf-ml-tasks-group-multimodal` `[ --count ]`
`./ai task --requiresCapability ibm-cap-contextual-understanding` `[ --count ]`
Evaluations
`./ai evaluation --byod` `[ --count ]`
`./ai evaluation --hasDocumentation arxiv.org/2310.12941` `[ --count ]`
`./ai evaluation ai_eval_PopQA` `[ --count ]`
`./ai evaluation --hasDataset truthfulqa/truthful_qa` `[ --count ]`
`./ai evaluation --hasTasks text-generation` `[ --count ]`
`./ai evaluation --hasLicense license-cc-by-4.0` `[ --count ]`
> Evaluations for Risks
`./ai evaluation --hasRelatedRisk atlas-hallucination` `[ --count ]`
`./ai evaluation --related --hasRelatedRisk mit-ai-causal-risk-timing-post-deployment` `[ --count ]`
Datasets
`./ai dataset --byod` `[ --count ]`
`./ai dataset CybersecurityBenchmarks_datasets_frr` `[ --count ]`
`./ai dataset --hasLicense license-apache-2.0` `[ --count ]`
`./ai dataset --hasDocumentation repo_nyu-mll_BBQ` `[ --count ]`
`./ai dataset --provider bigcode` `[ --count ]`
Adapters
`./ai adapter --byod` `[ --count ]`
`./ai adapter ibm-factuality-adapter-granite-3.2-5b-harm-correction` `[ --count ]`
`./ai adapter --hasDocumentation granite-guardian-paper` `[ --count ]`
`./ai adapter --hasAdapterType LORA` `[ --count ]`
`./ai adapter --implementsCapability ibm-cap-contextual-understanding` `[ --count ]`
`./ai adapter --adaptsModel granite-guardian-3.3-8b-instruct` `[ --count ]`
`./ai adapter --hasLicense license-apache-2.0` `[ --count ]`
> Adapting to Risk
`./ai adapter --hasRelatedRisk granite-relevance` `[ --count ]`
`./ai adapter --hasRelatedRisk granite-relevance --related` `[ --count ]`
LLMIntrinsics
`./ai intrinsic --byod` `[ --count ]`
`./ai intrinsic --hasDocumentation arxiv.org/2504.11704` `[ --count ]`
`./ai intrinsic ibm-factuality-intrinsic-jailbreak` `[ --count ]`
`./ai intrinsic --hasAdapter ibm-factuality-adapter-granite-3.3-8b-instruct-lora-citation-generation` `[ --count ]`
`./ai intrinsic --isDefinedByVocabulary ibm-factuality` `[ --count ]`
> LLMIntrinsics for Risks
`./ai intrinsic --hasRelatedRisk nist-confabulation` `[ --count ]`
`./ai intrinsic --related --hasRelatedRisk granite-answer-relevance` `[ --count ]`
Actions
`./ai action --byod` `[ --count ]`
`./ai action --isDefinedByTaxonomy nist-ai-rmf` `[ --count ]`
`./ai action --byod --isDefinedByTaxonomy acme-ai-taxonomy` `[ --count ]`
`./ai action acme-action-coffee-001 --byod` `[ --count ]`
`./ai action --hasAiActorTask 'Human Factors'` `[ --count ]`
> Actions for a Risk
`./ai action --hasRelatedRisk nist-human-ai-configuration` `[ --count ]`
`./ai action --related_ids --hasRelatedRisk atlas-toxic-output` `[ --count ]`
Controls
`./ai control --byod` `[ --count ]`
`./ai control --isDefinedByTaxonomy shieldgemma-taxonomy` `[ --count ]`
`./ai control gg-function-call-detection` `[ --count ]`
> Controls for Risk
`./ai control --detectsRiskConcept shieldgemma-dangerous-content` `[ --count ]`
`./ai control --hasRelatedRisk shieldgemma-hate-speech` `[ --count ]`
`./ai control --related --hasRelatedRisk atlas-toxic-output` `[ --count ]`
Incidents
`./ai incident --byod` `[ --count ]`
`./ai incident --isDefinedByTaxonomy ibm-risk-atlas` `[ --count ]`
`./ai incident ibm-risk-atlas-ri-fake-legal-cases` `[ --count ]`
> Incidents for Risks
`./ai incident --hasRelatedRisk atlas-dangerous-use` `[ --count ]`
`./ai incident --refersToRisk atlas-evasion-attack` `[ --count ]`
`./ai incident --hasRelatedRisk atlas-dangerous-use --related` `[ --count ]`
Documents
`./ai document --byod` `[ --count ]`
`./ai document repo_stanford_air_bench_2024` `[ --count ]`
`./ai document --hasLicense license-cc-by-4.0` `[ --count ]`
BenchmarkMetaCards
`./ai benchmarkcard` `[ --count ]`
LLM Question Policies
`./ai questionpolicy --byod` `[ --count ]`
Stakeholders
`./ai stakeholder --byod` `[ --count ]`
`./ai stakeholder --isDefinedByTaxonomy csiro-responsible-ai-patterns` `[ --count ]`
`./ai stakeholder csiro-stakeholder-ai-technology-producers` `[ --count ]`
`./ai stakeholder --isPartOf csiro-stakeholder-group-organization-level` `[ --count ]`
Organizations
`./ai organization --byod` `[ --count ]`
`./ai organization --grants_license license-cc-by-4.0` `[ --count ]`

Export cypher queries (Neo4J integration)
`./ai graph cypher --export --byod` `[ --count ]`
Export full Knowledge Graph
`./ai graph --export --byod` `[ --count ]`

And so on ..

${\color{green}Bring\ Your\ Own\ Data}$

Bring your own Data by adding schema-compliant yaml files to ./byo/data directory. See upstream readme.

This repository includes FINOS and related examples (ffiec, eu ai, iso42001, nist_sp_800_53, owasp_llm_t10, owasp_ml_t10, sr_11_7). Validate contributed LinkML Schemas as follows:

SDIR=".venv/lib/python3.14/site-packages/ai_atlas_nexus/ai_risk_ontology/"   # python version may vary
uv run linkml validate byo/data/*.yaml  -s ${SDIR}/ai-risk-ontology.yaml

${\color{green}Graph\ DB\ Visualization}$

Convert schema and instance data into Cypher representation to populate a Graph Database:

./ai graph cyper --export --byod

podman login dockerhub.registry.example.com
podman pull dockerhub.registry.example.com/neo4j
podman run --name ran_neo4j --rm --volume $(pwd)/graph/cypher:/examples --publish=7474:7474 --publish=7687:7687 --env NEO4J_AUTH=neo4j/demodemo  dockerhub.registry.example.com/neo4j:latest

Import into Neo4J:

docker exec --interactive --tty ran_neo4j cypher-shell -u neo4j -p demodemo
:source /examples/ai-risk-ontology.cypher

Open Neo4J Browser (login neo4j/demodemo) and CALL db.schema.visualization():

${\color{green}Crosswalk\ (Mappings)}$

Crosswalk documents are intended to provide a mapping of concepts and terms between the different taxonomies, frameworks, standards and regulation documents. The ./ai command produces a crosswalk between the risks in two taxonomies, finds related risks, and displays a subset of risk content in a pandas dataframe.

./ai crosswalk --isDefinedByTaxonomy nist-ai-rmf --isDefinedByTaxonomy2 finos-aigf --export --byod

${\color{green}CLI/API\ TESTING}$

Comprehensive unit tests are available for all CLI commands documented in this README.

101 test cases covering all CLI examples in both API and local modes
Automated server management for API mode tests
Parametrized tests for efficient coverage
CI/CD ready with GitHub Actions examples

# Verify test setup
uv run python lib/test/check_tests.py

# Run fast tests (API mode only, ~2 min)
,/scripts/tests.sh fast

See lib/test/README.md for detailed testing documentation.

LLM INFERENCING

Note: ${\color{orange}Requires\ LLM\ service!}$

Leverages Large Language Models (LLMs) to infer risk dimensions. You need access to LLM model (i.e. ibm-granite/granite-3.1-8b-instruct) and inference engines (vLLM) is required. On hosts with NVIDIA Container Toolkit, we can use the official vLLM container:

# example only
podman login dockerhub.registry.example.com
podman pull dockerhub.registry.example.com/vllm/vllm-openai

VLLM_API_KEY="${VLLM_API_KEY:-xxxxxxxxxxxxxxxxxxxxxxxxxyyyyyyyy}"
VLLM_HOST_IP="${VLLM_HOST_IP:-127.0.0.1}"
VLLM_API_URL="${VLLM_HOST_IP}/v1"
UV_TORCH_BACKEND=auto

podman run --device nvidia.com/gpu=all  -v ~/.cache/modelscope/hub/models:/root/.cache/modelscope/hub/models  -e REQUESTS_CA_BUNDLE=/cert.pem --mount type=bind,source=/etc/pki/tls/cert.pem,target=/cert.pem  --env "VLLM_USE_MODELSCOPE=True" --env "TRANSFORMERS_OFFLINE=1" --env "TORCH_CUDA_ARCH_LIST=8.6" --env "PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:64,expandable_segments:True" -p 8000:8000 --ipc=host  dockerhub.registry.example.com/vllm/vllm-openai   --model facebook/opt-125m --gpu_memory_utilization="0.8"

ARES Evaluation

ARES is an evaluation framework for Retrieval-Augmented Generation (RAG) systems. An extension for ai-atlas-nexus is available but a PR is needed for install.

ADOPT?

Open Data / Open Source / Inner Source promotes a community driven approach to curating and cataloguing resources such as datasets, benchmarks and mitigations.

${\color{green}Curate\ your\ own\ Open\ Data}$

Before curating the Model, its recommended you familiarize yourself with the basics of LinkML and its metamodel components. Like many modeling frameworks, LinkML comes with a controlled vocabulary of terms that are used to describe the model. While the modeling language is robust and might seem overwhelming, understanding just a few basic components is helpful.

${\color{green}Building\ Python\ Applications}$

Refer to the Python Reference. The SchemaView class in the linkml-runtime provides a method for dynamically introspecting and manipulating schemas. This can be used to programatically explore or edit the AI Atlas Nexus. See documentation and schemaview() function in ai.py.

./ai schemaview --introspect

Research

Pathways to Open Data, Linux Foundation

FilesExpand file tree

README.md

Latest commit

History