The knowledge base behind an AI calling agent, voice bot or WhatsApp chatbot in India

How to give an AI calling agent, voice bot or WhatsApp chatbot a knowledge base that holds up with Indian callers: Hindi and Hinglish, lakh and crore prices, GST, festive offers and thin websites.

An AI calling agent is only as good as what it reads. Whether you build it on Bolna, Vapi, Retell AI or ElevenLabs, and whether it answers calls or WhatsApp messages, it answers from a knowledge base: the business's services, fees, timings and rules. Indian businesses bring a few extra ways for that knowledge base to go wrong, and they are all fixable in how you write it.

This guide is for agencies and freelancers building voice bots and WhatsApp chatbots for Indian clients. Platform facts are from each platform's own documentation, checked on 25 September 2026. For how to load the file on each platform, see our platform setup guides.

Callers switch languages. The knowledge base shouldn't.

A caller to a Pune clinic may start in English, switch to Hindi for the question that matters and end in Marathi. It is tempting to write the knowledge base in all three. Don't: every fact then exists three times, and when the fee changes, one version gets missed.

Write the knowledge base in the language of the business's website, and let the agent answer in the caller's language. What you do need is search that works across languages, so a Hindi question finds an English fact:

  • Bolna lets you choose multilingual when you create a knowledge base, for cross-lingual retrieval "across 100+ languages". It can't be switched later, so choose it at the start.
  • ElevenLabs offers a multilingual embedding model among its RAG settings.
  • On any platform, test in the callers' languages: ask "Grade 6 ki fees kitni hai?" and check the agent found the Grade 6 fee, not another grade's.

Write every price the way a caller hears it

Indian prices come with more parts than most: lakh and crore, GST, per person, per cycle, per year, per sq ft. A voice agent reading "₹78L+GST" off a rate card can say anything.

  • One sentence per price, naming the product: "2 BHK apartments at Greenfield Heights start at 78 lakh rupees plus GST", not a row in a table.
  • Say the unit: per cycle, per month, per year, per person, per session.
  • Say the tax: "plus GST" or "including GST", in the same sentence.
  • Use the Indian number format the website uses (1,40,000, not 140,000), so the agent reads it the way callers expect.

Websites are thin. The details live elsewhere.

Many Indian small businesses have a five-page website, and the real details live in a PDF brochure, a rate card shared on WhatsApp, or the owner's head. A knowledge base built from the website alone is silent on exactly what callers ask: is parking included, is the bus fee extra, do you take EMI.

  • Add the documents: brochures, fee structures and rate cards, as files.
  • Write the unknowns down: "Whether the school bus fee is extra has not been confirmed; please call the office on +91 ..." The agent then says something true instead of guessing.
  • Ask the business once, with a list: collect every open question and send it to the owner in one go, rather than finding the gaps call by call.

Festive offers and seasonal prices

Diwali offers, admission-season discounts, monsoon rates and "limited period" deals appear on home pages and stay there long after they end. An agent that promises a festive offer from last year costs the business money or trust.

Write every offer with its dates, or as something to confirm: "The home page shows a festive offer of no stamp duty. The website gives no end date, so callers should confirm it with the sales office." Rebuild the knowledge base after each season.

Give the agent a phone number to fall back on

Indian callers expect to be handed to a person when the bot doesn't know. Write the right number into the knowledge base, in full with +91, next to each topic: admissions, the clinic's front desk, the site sales office. And add prompt lines that stop the agent inventing help:

Answer questions about the business only from the knowledge base. If it doesn't have the answer, say you will check, and give the business's phone number.
Give a fee, price, time or rule only from a sentence that names the service the caller asked about.
Whenever you give a price, say the name of the service with it.
If the caller hasn't said which service they mean, ask before answering.
When the knowledge base says something has not been confirmed, say exactly that. Never guess.
Don't promise WhatsApp messages, callbacks or bookings unless the knowledge base says the business offers them.

The last line comes from our own live calls: the agent promised a caller details on WhatsApp, and the business had no one to send them.

WhatsApp chatbots use the same file

A WhatsApp chatbot answers from its knowledge base exactly like a calling agent, and fails the same way: a fee without its course, a silence filled with a guess. Build one knowledge base per business and load it into both. When the website changes, rebuild it once and replace it in both.

A short test before the client goes live

Make these calls in the languages the business's callers actually use:

  1. Ask the fee or price of every service by name, one after another.
  2. Ask a follow-up in Hindi or Hinglish without repeating the service's name.
  3. Ask about two services with similar names (2 BHK and 2 BHK Premium, IUI and IVF, Grade 6 and Grade 7).
  4. Ask whether GST is included.
  5. Ask about the current offer.
  6. Ask something the website doesn't answer, and check the agent gives the phone number instead of guessing.

Where KB Builder fits

KB Builder reads the business's website and documents and writes the knowledge base in the website's own language: one sentence per fact, each naming its service, with the unit and tax, offers marked to confirm, and a list of what the website doesn't answer to send to the owner. It works for any business, from clinics and schools to real estate and resorts.

Common questions

What is an AI calling agent?

An AI calling agent is a voice bot that answers or makes phone calls for a business: it understands the caller, answers from the business's knowledge base, and books, qualifies or hands over. In India they are used by clinics, schools, coaching institutes, real estate developers, resorts and service businesses, often in Hindi or Hinglish as well as English.

Should the knowledge base be in Hindi if my callers speak Hindi?

Write it in the language the business's website uses, usually English, and let the agent answer in the caller's language. Translating the knowledge base adds a second version of every fact to keep in step. What matters more is that the platform's search works across languages: Bolna has a multilingual option you choose when creating the knowledge base, and ElevenLabs offers a multilingual embedding model.

Can the same knowledge base serve a voice bot and a WhatsApp chatbot?

Yes. A WhatsApp chatbot and a calling agent answer from a knowledge base in the same way and go wrong in the same way. One well-written file serves both.

How should prices be written for an Indian voice bot?

In full sentences that name the product, the amount, the unit and the tax: 'Grade 6 tuition at Riverside Academy is 1,40,000 rupees a year' or '2 BHK apartments start at 78 lakh rupees plus GST'. A price in a table or a bare '₹1.4L' can be misread or attached to the wrong product.

Sources

  1. Bolna docs: knowledge base (default and multilingual)
  2. Bolna API reference: create knowledge base
  3. ElevenLabs API reference: create agent (RAG embedding models)
  4. Retell AI docs: knowledge base