How does a document-grounded assistant use an office procedure while answering a question?

It answers only from general training without retrieving the instructions

It treats every stored document as equally relevant to every request

It retrieves relevant instructions and supplies them as context for the response

It must train a new foundation model from the beginning for each question

EKA-1 and Dhīr: How India’s New AI Models Connect Training, Language and Local Use

Why in news?

The IndiaAI Mission released three indigenous artificial-intelligence models in New Delhi on 3 November 2026: EKA-1, developed by Soket Labs Technology and Research, and Dhīr 2B and Dhīr 5B, developed by ZenteiQ AiTech Innovations. The companies received support under the mission’s IndiaAI Innovation Centre pillar, including subsidised computing infrastructure and assistance with research expenses.

What a foundation model provides

A foundation model is trained on a broad collection of data and can be adapted for different tasks. A language model learns patterns in text that help it generate responses. Developers can then adapt a common model for activities such as answering questions, writing code or working with a particular field’s documents.

This shared starting point matters because developing the underlying model requires substantial data, computing and technical expertise. Applications can build on that base while adding their own instructions, information sources and user-facing features. The foundation model and the complete application therefore occupy different parts of the development process.

Understanding parameters, tokens and inference

During training, a model adjusts numerical values called parameters as it learns from data. In language models, text is processed in smaller units called tokens. A token may represent a word, part of a word or a character, depending on the system used to divide the text.

The ministry states that EKA-1 has 26.3 billion parameters and was trained on approximately 7.5 trillion tokens. It supports all 22 scheduled Indian languages and English, with intended uses including mathematics, coding and reasoning. The two figures describe different features: the model’s learned numerical structure and the amount of text units processed during training.

Inference is the stage at which a trained model processes a user’s input and produces a response. Training establishes the model’s learned capabilities; inference puts those capabilities to work. This distinction helps explain how a model developed using large computing resources can later be deployed in an application.

Documents and local deployment

The official release describes Dhīr 2B and Dhīr 5B as open-source multilingual models supporting English and 10 Indian languages. Their intended uses include document-based automation, information retrieval and services in government, industry, education and defence. ZenteiQ describes them as capable of deployment on an organisation’s own premises and, on suitable devices, without continuous cloud connectivity.

Diagram shows a question and relevant retrieved documents supplied to a language model, which generates a response. — A document-grounded assistant supplies relevant source material along with the question. Clarity original conceptual diagram; sources: Lewis et al. (2020) and Google Machine Learning Glossary.

A document-grounded assistant uses relevant source material while preparing its response. For example, a system answering a question about an office procedure can first retrieve the relevant instructions and supply them to the language model. The response then draws on information provided for that particular question.

Running inference locally places the processing on the organisation’s or user’s hardware. This can support use where connectivity is limited and where documents need to stay within a controlled environment. The deployment arrangement connects the model with the equipment, documents and access controls of the organisation using it.

The wider IndiaAI Mission

The Union Cabinet approved the IndiaAI Mission on 7 March 2024. Its seven pillars cover computing capacity, indigenous models, datasets, application development, skills, startup finance, and safe and trusted AI. These address different needs in an AI system: the resources to develop models, the people and businesses that use them, and the methods for evaluating responsible deployment.

Conclusion

The launch brings multilingual model development together with public computing support and locally deployable applications. Understanding training, inference and document grounding makes the development easier to assess: each explains a different step between building a model and using it for a practical task.