Urdu chatbot projects are easiest to ship when they solve one narrow problem well: answering questions about a service, guiding students through a course, collecting support requests, or helping users find information. You do not need to train a large language model from scratch. A small intent-based bot, retrieval system, or hybrid assistant can deliver a reliable first version with modest data and infrastructure.
For Indian builders, the hard parts are usually not only model selection. They include Urdu script, Roman Urdu, code-switching with Hindi and English, spelling variation, privacy, and evaluation with real users. This guide explains how to build a practical small Urdu chatbot as of 2026.
1. Define a narrow job
Start with a written scope rather than an open-ended “talk to users” objective. Specify:
- Audience: for example, Urdu-speaking customers in Delhi, Hyderabad, Uttar Pradesh, or Jammu and Kashmir.
- Channel: website, WhatsApp, Telegram, mobile app, or an internal support tool.
- Tasks: answer FAQs, check an application status, recommend learning material, or create a support ticket.
- Escalation rule: identify when the bot must hand the conversation to a person.
- Success metric: task completion, correct answers, resolution rate, or reduced response time.
A focused bot needs fewer examples and is easier to test. If your project must support multiple Indian languages, review building AI apps for the next billion users in India before deciding on your interface, latency, and data strategy.
2. Choose the right architecture
There are three sensible starting points.
Rule-based bot
Use fixed intents, keywords, and response templates. This is appropriate for menus, form collection, and high-stakes workflows where predictable behaviour matters more than flexible conversation.
Retrieval-based bot
Store approved Urdu documents, split them into passages, retrieve relevant passages for each question, and generate or return an answer grounded in those passages. This works well for schemes, product manuals, school content, and help centres that change over time.
Hybrid bot
Use intent classification for actions and retrieval for informational questions. A small language model can rewrite or classify input, while business logic controls what the bot is allowed to do. For most first products, this is the best balance between quality, cost, and control.
For language and dataset decisions, the guide to low-resource Indic natural language processing is especially relevant. Urdu has fewer labelled resources than English, so careful data curation often matters more than choosing a larger model.
3. Handle Urdu input correctly
Urdu uses the Arabic-derived Nastaliq writing system, but users may type in different forms. Your input pipeline should preserve the original message and create a normalised copy for search and classification.
Support at least these patterns:
- Urdu script, such as
مجھے درخواست کی معلومات چاہیے. - Roman Urdu, such as
mujhe darkhwast ki maloomat chahiye. - Mixed Urdu-English or Urdu-Hindi messages.
- Different Unicode forms, punctuation, whitespace, and spelling variants.
- Informal abbreviations, typos, and speech-to-text errors.
Do not remove diacritics, punctuation, or characters blindly. Normalisation should be reversible and tested against real examples. Keep language detection probabilistic: a short message such as “status check karna hai” may be mixed rather than purely Urdu.
A useful data record stores the raw text, normalised text, detected language, intent, entities, expected response, and whether escalation is required. Keep user identifiers separate from conversation content wherever possible.
4. Build a small, representative dataset
Begin with 10–20 intents and 20–50 examples per intent, then expand based on failures. Include genuine variation rather than repeating the same sentence with minor edits.
For every intent, collect:
- Formal Urdu and conversational Urdu.
- Roman Urdu alternatives.
- Code-switched queries with English terms.
- Misspellings and incomplete messages.
- Similar questions that belong to different intents.
- Out-of-scope and adversarial requests.
For example, a “check application status” intent might include questions about a submitted form, pending verification, rejected documents, and a status link. Label those separately if the system takes different actions.
Ask native or highly proficient Urdu speakers to review translations and responses. Machine translation can accelerate drafting, but it should not be the final quality check for tone, politeness, names, dates, or government terminology.
5. Implement a baseline in Python
A simple production-minded flow can be built with Python, a web framework such as FastAPI, and a model or search service:
1. Receive the message and assign a request ID.
2. Validate length, remove unsafe control characters, and retain the raw input.
3. Detect script and language mix.
4. Normalise a copy for intent classification and retrieval.
5. Classify the intent and extract entities such as dates, locations, or application numbers.
6. Route transactional intents to deterministic code.
7. Retrieve approved content for informational questions.
8. Generate a response only within the retrieved context and policy.
9. Return Urdu text, a clarifying question, or a human handoff.
10. Log the outcome without storing unnecessary personal data.
Keep prompts and response templates versioned. Add confidence thresholds: a low-confidence result should ask one precise clarifying question or escalate, not invent an answer. For sensitive domains such as health, finance, legal services, or welfare applications, prefer verified content and human review.
6. Evaluate more than fluency
A chatbot can sound natural and still fail users. Build a test set that was not used for development and measure:
- Intent accuracy: whether the correct workflow was selected.
- Entity accuracy: whether names, dates, numbers, and locations were extracted correctly.
- Groundedness: whether answers are supported by approved sources.
- Task completion: whether users reached the intended outcome.
- Abstention quality: whether the bot refuses or escalates appropriately.
- Latency and cost: especially on mobile networks and low-end devices.
- Script coverage: Urdu, Roman Urdu, and mixed-language performance.
Review errors by category. Common failures include confusing similar intents, mishandling Roman Urdu, losing digits in application numbers, and returning Hindi vocabulary when the user expects Urdu. A small native-speaker review panel is often more valuable than an automated fluency score.
7. Make deployment affordable and safe
For a pilot, run the API in a container on a modest cloud instance or Indian cloud region, use a managed database, and cache static answers. Keep retrieval indexes and model services replaceable so that you can change providers without rewriting the product.
Use HTTPS, encrypted secrets, access controls, rate limits, abuse monitoring, and regular backups. Avoid sending personally identifiable information to an external model unless users have consented and the arrangement is appropriate. Define retention periods before launch.
If the chatbot eventually needs speech input or phone support, treat that as a separate system. Review how to build a voice agent and compare it with text-first designs through voice agent vs chatbot. Speech introduces recognition errors, accent variation, latency, and additional consent requirements.
8. Launch in controlled stages
Ship a private alpha with staff or trusted users, then release to a small percentage of traffic. Add a visible “talk to a person” option and collect feedback in Urdu as well as English. Review failed conversations weekly, update the dataset and knowledge base, and re-run the fixed evaluation set before every release.
Do not optimise for long conversations. A useful small chatbot may answer one question, complete one form, and stop. That is a stronger product outcome than an entertaining bot that gives unreliable information.
FAQ
Do I need to train my own Urdu model?
Usually not. Start with an existing multilingual model, retrieval, and a carefully labelled dataset. Fine-tune only after you can identify repeatable errors and have enough licensed examples.
Should I support Roman Urdu?
Yes, if your users type on mobile keyboards or use messaging platforms. Treat Roman Urdu as a first-class input format rather than forcing users to switch scripts.
What is the cheapest useful MVP?
A narrow FAQ or intent bot with approved responses, a confidence threshold, a human fallback, and basic analytics. Add generative features only where they improve a measured task.
Where can Indian AI teams find support?
Document the problem, dataset permissions, evaluation plan, and expected users before seeking funding. AI Grants India can help founders and researchers discover relevant grant opportunities.