How to Make AI More Accurate: An Indian Perspective
By ICTpost Research Desk | 3 July 2025
“AI is never wrong, right?”
Well… not really.
In a country where AI is rapidly entering schools, panchayat offices, and digital public infrastructure, accuracy isn’t just a feature — it’s a necessity.
From predicting crop yields in Telangana to powering chatbots for government schemes like PM-Kisan and Ayushman Bharat, AI is becoming deeply woven into India’s social and governance fabric. But unlike humans, AI systems sometimes make mistakes that are subtle, confident, and dangerously wrong. This is especially concerning in high-stakes use cases involving health, law, or financial entitlements.
The latest ICTpost report found that 71% of Indian enterprises using AI had “accuracy of AI predictions” as their top concern. In another study by IIT Madras, an AI chatbot trained on legal data produced hallucinated case citations in 34% of queries unless fine-tuned or externally augmented.
So how do we fix this?
Let’s break down the top 8 strategies to improve AI accuracy, and how each one can be applied effectively in India.
1. RAG (Retrieval-Augmented Generation): Bring in Real-Time Knowledge
What is it?
RAG models improve response accuracy by retrieving real-time information from a trusted source (e.g., PDFs, government documents, news updates) before generating an answer.
Why it matters in India:
Most AI models are trained on data that ends in 2023 or earlier. If you ask about new UGC regulations, changing farm subsidies, or state-specific caste certificates, a basic AI model may hallucinate — or worse, fabricate.
Example:
In 2024, NIC piloted a chatbot for PM Kisan Yojana FAQs using RAG architecture. Accuracy of responses improved from 63% to 92% after adding real-time retrieval from the Agri Ministry’s database.
Toolkits: LangChain + FAISS + ChatGPT APIs
Use Case: VLE chatbot answering state-specific scheme queries with real documents.
2. Choose the Right Model for the Right Task
Not every job needs GPT-4. Smaller, domain-specific models are often more accurate and cost-effective for specific tasks.
A Bengaluru startup, LegalMind AI, tested a general-purpose LLM vs. a fine-tuned Indian legal model (based on Supreme Court archives).
Result? The fine-tuned legal model achieved 89% citation accuracy, while the general model was at 64% — despite being larger.
Use general models (like GPT-4 or Claude) for broader tasks.
Use fine-tuned models (like IndicBERT, BharatLLM, or Bhashini-powered AIs) for Indian languages or niche sectors.
3. Chain of Thought (COT) Prompting: Make the Model Show Its Work
Instead of jumping to conclusions, the AI is asked to reason step-by-step. This dramatically reduces logic errors — especially in math, decision trees, or entitlements.
Example:
CSC-SPV used COT prompting in an internal AI system to compute Ayushman Bharat eligibility.
Without reasoning prompts, the model misclassified 22% of BPL family cases.
With COT, errors fell to under 6%.
Implementation: Add phrases like “Let’s think step-by-step” or use few-shot examples with correct reasoning.
4. LLM Chaining: Ask More Than One Brain
Instead of relying on one model, pass the same query through multiple models, and either vote or use a supervisor to pick the best answer. In an AI-based legal advice tool tested in Delhi Legal Aid Centers (2024), using three LLMs in a chained architecture increased correct outputs by 31% compared to relying on a single model. AI for medical advice, RTI drafting, or court form filling should never rely on a single model. Consensus is safer.
- Mixture of Experts (MoE): Let the Specialists Handle It
An MoE architecture consists of sub-models specialized in different domains (health, finance, language). A “gatekeeper” routes your query to the right expert(s).
Think of an AI hospital where there’s a cardiologist model, a neurologist model, and a general physician model — your query is sent to the right doctor. Example: Google’s Switch Transformer, an MoE model, used expert routing to achieve 7x faster training and better accuracy with fewer resources.
- Temperature Settings: Accuracy vs Creativity
What is it?
AI models use a temperature setting between 0 and 1 to control response randomness.
0.0 – 0.3: Factual, consistent (best for legal, medical, technical queries)
0.7 – 1.0: Creative, diverse (best for slogans, stories, poetry)
Example:
When used for generating Panchayat slogans, a temperature of 0.8 produced more engaging lines.
But for exam result queries, reducing temperature to 0.2 improved factual consistency.
Tip:
Tune temperature per task in your app or API call.
- System Prompts: Set Ground Rules
A system prompt is an instruction embedded before user queries, telling the model how to behave.
Example: For a CSC chatbot, a system prompt could be:
“Only answer using data from CSC-Gov.in. Do not guess. Always ask for location if district-specific information is required.”
Impact:
System prompts can reduce hallucinations by up to 40%, according to OpenAI documentation.
- Reinforcement Learning with Human Feedback (RLHF)
Humans rate AI responses — good or bad. These ratings are then used to fine-tune future responses.
Indian Context: Deploying AI at scale (e.g., IndiaAI’s National Language Translation Mission) will only succeed if millions of teachers, VLEs, and users rate answers over time.
In a Telangana EdTech pilot, student ratings were used to fine-tune an AI tutor in Telugu and English.
After 10,000 ratings, test scores improved by 18% among users.
In Summary: Towards a More Accurate, Bharat-Ready AI
| Technique | Impact | Example |
|---|---|---|
| RAG | +25-40% accuracy | PM Kisan chatbot, Agri schemes |
| Right Model | +20% | LegalMind AI vs general LLM |
| COT Prompting | -75% logic errors | Eligibility calculators |
| LLM Chaining | +30% correctness | Legal/RTI support |
| Mixture of Experts | +Efficiency | Health & Governance routing |
| Temperature Tuning | +Consistency | Exam result vs poetry prompts |
| System Prompts | -Hallucinations | Scheme-aware chatbots |
| RLHF | +Human alignment | Telugu AI tutor |
The path to trustworthy AI in India isn’t about one big breakthrough — it’s about thousands of small, accurate answers — given at the right place, in the right language, and with the right logic.
Tell Us What You Think
Have you deployed or tested AI in an Indian context? What methods work for you?
Write to us at editor@ictpost.com or tag us on social media with #AccurateAIIndia
Let’s build the future of Indian AI — not just powerful, but precise.
Because in Bharat, accuracy is empowerment.
