A language model that has read everything and remembers nothing exactly is a poor witness. It will answer a question about a classical Ayurvedic formulation fluently, confidently, and — often enough to matter — wrongly, because it is reconstructing a plausible answer from statistical patterns rather than reading a text.
Medicine is the worst possible place for that failure mode. On 19 September 2026 the Ministry of Ayush set out how it is trying to avoid it.
The architecture that does the work
The centrepiece is a use case titled "AI-driven Decision Support for Ayurveda: A Retrieval-Augmented Generation (RAG) Approach", which featured in the AI Impact Casebook on Health developed in partnership with the World Health Organization and presented at the IndiaAI Impact Summit 2026.
Retrieval-Augmented Generation is worth understanding properly, because it is now the dominant architecture for serious applications of language models and appears increasingly in questions about AI.
A plain large language model answers from its parameters — the weights adjusted during training. It has no access to a document at the moment you ask; it produces what its training makes statistically likely. When it does not know, it does not stop. It generates something that reads like an answer. This failure is called hallucination, and its distinguishing feature is confidence: a hallucinated answer is not hesitant or garbled, it is fluent and wrong.
A RAG system inserts a step before generation. When a question arrives, the system first retrieves relevant passages from a defined corpus — here, classical Ayurvedic literature and research material — and then asks the model to generate an answer grounded in those retrieved passages.
Three consequences follow, and they are the reason the architecture exists:
Grounding. The answer is constrained by documents actually fetched, sharply reducing fabrication.
Traceability. Because the retrieved passages are known, the answer can be cited. In a clinical context, an answer nobody can trace to a source is not usable, however correct it happens to be.
Currency without retraining. The corpus can be updated at any time. A model cannot be retrained every time a new paper appears; a retrieval index can be refreshed overnight.
For a knowledge system built on a large, fixed and textually precise classical literature, the fit is close to ideal. Ayurveda's authority rests on specific texts. RAG is the architecture that makes a model quote them instead of paraphrasing from memory.
The applications on the Ayush Grid
The Ministry describes this as building on Ayush Grid, its digital backbone, and lists six AI assistants:
- MAISP Chatbot — a conversational interface for Ayush information, including Ayurveda
- ARP Research Assistant — a GenAI assistant extending the Ayush Research Portal, helping researchers navigate the literature
- eLMS Chatbot — AI-based student life-cycle management for Ayush education institutions
- e-CHARAK Chatbot — a medicinal-plant trade assistant, connecting conversational AI to the supply ecosystem behind Ayurvedic pharmacology
- Yoga Saarthi — a yoga recommendation assistant
- Nivesh Saarthi — an investment assistant for the Ayush sector
The list is more interesting than it first appears because of what it covers: a citizen-facing service, a research tool, an education administration system, a commodity trade platform, a wellness recommender and an investment facilitator. Only two of the six are about health advice. The rest are about running a sector — and that is a fair picture of where AI actually delivers value in government, which is rarely the dramatic clinical use case and usually the unglamorous administrative one.
Languages, and why 22 is the operative number
Through collaboration with IndiaAI and the Digital India Bhashini Division (DIBD) under MeitY, Ayush Grid portals are now available in the 22 scheduled languages of India. Ministry artefacts have also been placed on AI Kosh to support wider AI research in this domain.
The phrase 22 scheduled languages refers to the languages listed in the Eighth Schedule of the Constitution — a fact worth carrying, since it converts a technology story into a polity answer.
The reason multilingual access matters more here than in most digital services is specific. Traditional medicine is used most heavily by people who are least likely to operate comfortably in English, and a health information service available only in English reaches the population that least needs it. Bhashini, India's national language technology mission, exists to make translation and speech infrastructure available across government services, and this is a clean example of why.
There is a genuine technical difficulty behind the announcement that is worth naming. Language models perform markedly better in languages with abundant digital text. Most Indian languages are low-resource by that measure — the quantity of digitised, high-quality text in Maithili or Santali is a tiny fraction of what exists in English. Building usable models for them is not a matter of running the same process on a smaller corpus; it requires deliberate data creation, which is much of what Bhashini and AI Kosh are for.
The line the Ministry draws
The Ministry states the intent plainly: not to replace traditional knowledge with technology, but to use AI to make that knowledge more discoverable, evidence-based and accessible.
That framing deserves engagement rather than acceptance, because it sits on a real tension.
AI applied to a traditional knowledge system can do two quite different things. It can make existing knowledge easier to find and apply — searching a large classical corpus, matching a described presentation to documented formulations, translating a text. Or it can be used to make claims of efficacy that the underlying evidence does not support, with the model's fluency supplying an authority the data does not.
The first is straightforwardly useful. The second is the risk, and it is the same risk that runs through the regulation of traditional medicine generally — a subject examined this week in our piece on the Supreme Court's limits on stem cell therapy, where the line between what a therapy is documented to do and what it is marketed as doing was exactly what the Court enforced.
The WHO, ITU and WIPO addressed this jointly in a technical brief released on 11 July 2025, Mapping the Application of Artificial Intelligence in Traditional Medicine, which examined how AI can support traditional medicine while stressing responsible, evidence-informed and culturally appropriate approaches. That three-agency combination is itself instructive: WHO for health, ITU for the technology standards, WIPO for the intellectual property questions that arise the moment traditional knowledge is digitised — the biopiracy problem that India's Traditional Knowledge Digital Library was built to address.
The Ministry's stated direction — use cases in voice-based Electronic Health Records, disease-trend detection and personalised treatment support, pursued through the IndiaAI Innovation Challenge — sits mostly on the useful side of that line. Voice-based EHR in particular is a sensible target: the reason Indian clinical records are poor is not that doctors are unwilling but that typing during a consultation is slow, and a practitioner seeing sixty patients a day will not do it.
🔑 Revision block
The announcement. 19 September 2026, Ministry of Ayush — AI programme built on Ayush Grid, the Ministry's digital backbone.
RAG — Retrieval-Augmented Generation. The system retrieves relevant passages from a defined corpus first, then generates an answer grounded in them. Gains: grounding (less fabrication), traceability (the answer can be cited), and currency without retraining (refresh the index, not the model).
Hallucination. A language model producing a fluent and confident but false answer from its parameters. Fluency, not hesitancy, is what makes it dangerous.
The showcase. "AI-driven Decision Support for Ayurveda: A Retrieval-Augmented Generation (RAG) Approach" — featured in the AI Impact Casebook on Health, developed with the World Health Organization, at the IndiaAI Impact Summit 2026.
Six assistants. MAISP Chatbot (citizen information) · ARP Research Assistant (Ayush Research Portal, GenAI) · eLMS Chatbot (student life-cycle management) · e-CHARAK Chatbot (medicinal-plant trade) · Yoga Saarthi (yoga recommendation) · Nivesh Saarthi (investment facilitation).
Language layer. Collaboration with IndiaAI and the Digital India Bhashini Division (DIBD), MeitY. Ayush Grid portals now in the 22 scheduled languages — those in the Eighth Schedule of the Constitution. Ministry artefacts placed on AI Kosh.
Low-resource languages. Models perform worse where digitised text is scarce; most Indian languages are low-resource, which is why Bhashini and AI Kosh focus on data creation, not just model training.
The international frame. WHO, ITU and WIPO technical brief, 11 July 2025 — Mapping the Application of Artificial Intelligence in Traditional Medicine; stresses responsible, evidence-informed and culturally appropriate approaches. The three agencies map to health, technology standards and intellectual property.
Planned use cases. Voice-based Electronic Health Records · disease-trend detection · personalised treatment support, via the IndiaAI Innovation Challenge.
The tension to state honestly. AI can make existing knowledge easier to find and apply, or it can lend unearned authority to efficacy claims the evidence does not support. The first is the stated intent; the second is the risk.
🎯 Practice MCQs
Q1. Retrieval-Augmented Generation refers to an approach in which a model: (a) Is retrained on new data before every query (b) Retrieves relevant documents and generates an answer grounded in them (c) Generates multiple answers and selects the longest (d) Translates a query into English before answering
→ (b)
Q2. "Hallucination", in the context of large language models, means: (a) The model failing to produce any output (b) The model producing a fluent but factually false answer (c) The model refusing to answer a sensitive question (d) The model producing output in the wrong language
→ (b)
Q3. A principal advantage of RAG over a plain language model in a clinical setting is that: (a) It runs faster (b) Its answers can be traced to a source document (c) It requires no computing hardware (d) It works without any training data
→ (b)
Q4. The 22 scheduled languages of India are listed in which Schedule of the Constitution? (a) Sixth Schedule (b) Seventh Schedule (c) Eighth Schedule (d) Ninth Schedule
→ (c)
Q5. Bhashini is associated with: (a) Satellite navigation (b) Language technology and translation for government services (c) Semiconductor fabrication (d) Genomic sequencing
→ (b) — under MeitY.
Q6. The technical brief Mapping the Application of Artificial Intelligence in Traditional Medicine was released jointly by: (a) WHO, ITU and WIPO (b) UNESCO, WHO and FAO (c) UNDP, WTO and WIPO (d) WHO, UNEP and ITU
→ (a) — health, technology standards and intellectual property respectively.
Q7. In the context of AI, a "low-resource language" is one that: (a) Is spoken by few people (b) Has limited digitised text available for training (c) Lacks a written script (d) Is not officially recognised
→ (b)
Q8. e-CHARAK, referred to in the Ministry's AI programme, is associated with: (a) Medicinal plant trade (b) Yoga instruction (c) Hospital management (d) Drug licensing
→ (a)
Q9. A key advantage of updating a retrieval corpus rather than retraining a model is that: (a) It eliminates the need for any model (b) New information can be incorporated without the cost of retraining (c) It removes the possibility of error (d) It reduces the number of languages supported
→ (b)
Q10. Consider the following statements: 1. A plain large language model answers from documents fetched at the time of the query. 2. The Ayush AI use case featured in a health casebook developed in partnership with the WHO. Which is/are correct? (a) 1 only (b) 2 only (c) Both 1 and 2 (d) Neither 1 nor 2
→ (b) — a plain model answers from its trained parameters; fetching documents at query time is what RAG adds.
📋 How this gets asked (PYQ pattern)
Artificial intelligence has moved from an occasional curiosity in the NDA General Knowledge paper to a recurring topic, and the questions have become more technical as the subject has become more familiar.
The terminology question now goes beyond "what is AI". Terms such as machine learning, deep learning, neural network, generative AI, large language model, training data, hallucination and now RAG each have a one-sentence definition worth having ready. The nesting is worth learning as a hierarchy: artificial intelligence contains machine learning, which contains deep learning, which is the basis of modern large language models.
The initiative question asks which body runs what — IndiaAI Mission and AI Kosh under MeitY, Bhashini for language technology, the IndiaAI Impact Summit as the flagship event. Mapping initiative to ministry is the same preparation habit that pays across schemes.
The polity crossover is where the efficient marks are. A technology story mentioning "22 scheduled languages" is one step from an Eighth Schedule question, and candidates who spot that link answer a polity question from a science article. This happens often enough to be worth watching for deliberately.
The international body question attaches an organisation to a domain — WHO to health, ITU to telecommunications and technology standards, WIPO to intellectual property. A three-agency brief is an unusually convenient way to fix all three in memory at once.
Preparing for the NDA? Technology questions reward precise one-line definitions over general familiarity — knowing what RAG adds to a language model is worth more than knowing that AI is important. Build the base with our NDA study material on science, follow the daily NDA current affairs, and prepare with our faculty in the upcoming Cavalier courses in Delhi.
✍️ Written by Col Vijyanat Thakur — Defence studies & general studies faculty at The Cavalier. Reviewed by the Cavalier Faculty Desk.