Author:
- Susan Malaika, Senior Technical Staff Member at IBM
- David Manset, Senior Advisor Open Source and Digital Public Infrastructure
- Roman Chestnov, Project Officer Digital Public Services
Quick answer: ITU’s GENIE.AI Framework is an open source generative AI platform for public sector institutions, built on the LF AI & Data projects OPEA (Open Platform for Enterprise AI) and Docling, alongside IBM Granite models.
It enables governments to move from AI experimentation to practical public-sector services, including chatbots, digital assistants and knowledge assistants, without vendor lock-in, while building local technical capacity and reusable digital capabilities.
Key takeaways
- GENIE.AI is a fully open source, DPG compliant, GovStack-aligned generative AI framework built for public institutions with limited budget and technical capacity.
- It’s leveraging OPEA, an LF AI & Data project, which provides the retrieval-augmented generation (RAG) orchestration backbone.
- Docling, also an LF AI & Data project, strengthens document ingestion by improving extraction accuracy from complex PDFs, a common challenge with government records.
- IBM Granite serves as the main generation model, run through vLLM.
- The framework powers pilot deployments from the ITU-IEEE GenAI for Good Challenge, in Lesotho, the Gambia, and Bangladesh.
- GENIE.AI has been released publicly under the Apache 2.0 license via the United Nations’ Open Source United GitLab.
The framework also aligns with the direction set by ITU Member States through WTDC-25 Resolution 91, which calls for strengthening AI readiness, infrastructure and capacity across developing countries. GENIE.AI represents one practical approach to translating that mandate into deployable public-sector solutions.
Why Public Sector AI Needs an Open Source Framework
Across the globe, public institutions are eager to harness generative AI to better serve their citizens, but many face steep barriers: limited budgets and infrastructure, thin technical capacity, and a dependence on proprietary vendors that can undermine long-term sovereignty over critical systems.
Public institutions need more than access to AI models, they need the ability to adapt, govern and sustain AI solutions according to national priorities. GENIE.AI was designed to help close that capability gap.
What Is the GENIE.AI Framework?
GENIE.AI is a fully open source, modular platform that lets public sector organizations design, deploy, and manage their own generative AI solutions, including chatbots, digital assistants, and content-generation tools, at low cost and without vendor lock-in. It’s aligned with GovStack digital public infrastructure standards, and the Digital Public Goods Alliance (DPGA) standard on Digital Public Goods, so it can plug directly into the broader ecosystem of government digital services, and it’s designed from the ground up to be dockerized, Kubernetes-ready, and community maintained.
Model choice is part of that openness: GENIE.AI is model-agnostic by design, enabling governments to select and integrate the open-source LLM best suited to their needs, resources and national context. Furthermore, the framework is optimized for smaller LLMs, making it suitable for edge deployment and application in resource-constrained environments. Enhancements and optimizations introduced to the core RAG microservices (guided by public-sector requirements), combined with tunable wrappers, help compensate for the performance difference between smaller and frontier LLMs, enabling reliable, high-quality output even with less capable models.
Additionally, rather than positioning AI as a standalone technology, GENIE.AI explores how open-source AI can complement Digital Public Infrastructure and Digital Public Goods, allowing governments to build intelligent public services on top of reusable digital foundations.
The framework grew out of an ITU-led informal multi-stakeholder Working Group on Open-Source Generative AI for Public Services, and it is built with various open-source stacks including the OPEA, giving GENIE.AI a proven, production-oriented foundation for its retrieval-augmented generation (RAG) pipeline rather than a from-scratch architecture.
Where OPEA and Docling Fit Into the Architecture
GENIE.AI leverages OPEA’s orchestration framework for RAG and introduces specialized versions of core microservices, enhanced with custom logic to maximize performance, increase tunability, and adapt to common public-sector requirements. A modular architecture, combined with the OPEA service registry for orchestration, allows governments to assemble AI applications that remain portable and interoperable. Additionally, enhancements to traceability and observability provide greater control, which is particularly important in the public sector. .
- ChatQnA backend orchestrator: an extension of OPEA ChatQnA, this is the central controller that defines the execution graph across services, handles optional translation and guardrail steps, manages chat history, and returns a structured response (generated text plus metadata like confidence scores and source documents) back to the frontend.
- Dataprep service: a custom version of OPEA dataprep microservice, this component ingests documents (PDF, DOCX, XLSX, TXT, MD, HTML), chunks and labels them, and extracts entities and relationships to build a knowledge graph. This is where Docling comes in: GENIE.AI incorporates Docling as a default text-extraction pipeline specifically to improve the accuracy and resilience of parsing complex PDF documents, a common pain point for government records, policy documents, and forms.
- Retriever service: extends OPEA’s retriever microservice with ArangoDB integration and a layered labelling system, combining vector similarity search, keyword search, label-based filtering, and knowledge graph traversal for hybrid retrieval.
- Embedding and reranker services: served via Text Embedding Inference (TEI), using swappable open models and configurable wrappers to better control system behavior across different types of content.
- LLM services: powered by vLLM for inference, with IBM Granite (granite-4.1-8b) serving as the main generation model and Google’s Gemma handling translation and guardrail duties.
In short, OPEA supplies the architectural backbone and orchestration pattern for the whole RAG pipeline, Docling strengthens the document-ingestion layer where public sector data is often messiest, and Granite provides the generation engine, together forming a fully open stack that public institutions can run, audit, and extend themselves.
Real-World Impact: The GenAI for Good Challenge
This isn’t just an architecture diagram, it’s already being deployed. Earlier this year, ITU and IEEE HT ran the first GenAI for Good Challenge, inviting teams from around the world to build solutions on the GENIE.AI framework, powered by Granite,Docling and additional open source LLM models, for real government use cases in health, agriculture, and climate. More importantly, the Challenge demonstrated that an open-source AI framework could become a shared innovation platform connecting governments, universities, developers and international organizations around common public-sector challenges. From more than 300 submissions across 79 countries, three teams were selected as winners and announced at AI for Good Week in Geneva:
- AgriConnect Lesotho: Keletso, an AI-powered assistant giving Lesotho’s smallholder farmers localized guidance on planting, pests, and harvesting.
- Innov8AI (The Gambia): ASante, a digital wayfinder helping people turn trusted health information into concrete next steps for chronic disease prevention.
- Polisense AI (Bangladesh): MEWA, an early weather advisor translating complex weather data into clear Bengali-language guidance for farmers.
All three teams are now entering a three-month deployment and scale-up phase in collaboration with the governments of Lesotho, the Gambia, and Bangladesh, a concrete example of open source infrastructure translating into real government services. A second edition of the Challenge, focused on building better cities, is launching this fall and will again build on GENIE.AI, OPEA, and Docling.
Additionally, in El Salvador, FAO and ITU are collaborating with the National Centre for Agricultural and Forestry Technology (CENTA) on piloting a GENIE.AI-based assistant for agriculture extension workers providing guidance to farmers in remote rural areas.
A Growing Community Conversation
During 2026, GENIE.AI evolved from a technical prototype into a growing international community of practice spanning governments, UN agencies, academia, open-source communities and technology partners. At Community Day on June 26 at One Madison, David Manset, who, among his many roles, serves as Vice Chair of the Open Source United Nations Community of Practice, presented GENIE.AI in the SDGs track, with Red Hat’s Hannah Braswell attending and recording the session. Lauren McHugh also presented the framework at a meetup hosted at St. John’s University, and GENIE.AI featured prominently at AI for Good Week in Geneva on July 10, in a session on building trustworthy, open source foundations for public sector generative AI.
Open source is not only about reducing costs. It is about giving governments the freedom to understand, adapt and build AI according to their own priorities,” said David Manset, ITU Telecommunication Development Bureau.
What’s Next for GENIE.AI
The GENIE.AI team released the framework earlier this year via an official UN GitLab repository under the Apache 2.0 license, complete with documentation and deployment guides. The roadmap ahead includes expanded agentic capabilities and integrations with custom tools, a lightweight speech-to-text and text-to-speech support, social messaging platform integrations, and expanded database and Kubernetes deployment options, all developed in the open, with community contribution actively encouraged.
For the LF AI & Data community, GENIE.AI demonstrates how mature open-source AI projects can create impact far beyond enterprise environments. By combining OPEA’s orchestration, Docling’s document intelligence and Granite’s language capabilities within an open public-sector framework, the partnership illustrates a broader vision: helping governments develop the capacity to build, adapt and govern AI solutions that remain open, reusable and locally sustainable.
The opportunity is not simply to deploy better AI. It is to make high-quality public-sector AI increasingly accessible to every country, regardless of its starting point.
Open-source AI technologies become far more valuable when they are assembled into reusable public digital capabilities that governments can own, adapt and scale.
Frequently Asked Questions
What is the ITU GENIE.AI Framework? GENIE.AI is an open source, DPGA and GovStack-aligned generative AI framework that lets public sector institutions build and deploy chatbots, digital assistants, and RAG-based tools at low cost, without dependence on a single vendor.
What is OPEA and how is it used in GENIE.AI? OPEA (Open Platform for Enterprise AI) is an LF AI & Data project that provides the orchestration framework for GENIE.AI’s backend. Its ChatQnA megaservice template coordinates the retrieval-augmented generation pipeline, including the dataprep, retriever, embedding, reranker, and LLM services.
What is Docling and what role does it play? Docling is an LF AI & Data project for document parsing and extraction. Within GENIE.AI, it’s used in the dataprep service as a default pipeline to improve the accuracy and resilience of text extraction from complex PDF documents, which are common in government records and forms.
Which LLM does GENIE.AI use? GENIE.AI is model-agnostic by design, allowing governments and developers to select and integrate the open-source LLM best suited to their requirements, resources and use cases. The current reference implementation demonstrates this flexibility using IBM Granite (granite-4.1-8b) as its main generation model, served through vLLM, with Google’s Gemma supporting translation and guardrail tasks.
Is GENIE.AI open source? Yes. GENIE.AI is fully open source and will be publicly released under the Apache 2.0 license via an official UN GitLab repository later this year.
Where has GENIE.AI been deployed? GENIE.AI has powered winning solutions from the ITU-IEEE GenAI for Good Challenge, now being deployed and scaled in collaboration with the governments of Lesotho, the Gambia, and Bangladesh, covering agriculture, health, and climate use cases. Additionally, ITU is collaborating with the FAO and the government of El Salvador on piloting a GENIE.AI-based solution for an agriculture extension call center.
