Multilingual RAG Chatbot Case Study: How We Built NovaMind AI, a Voice and Text Assistant for Customers Who Switch Languages

Written by Anuj Kumawat, AI/ML Engineer, 5+ years in machine learning and applied AI. Worked on NovaMind’s retrieval and language pipeline.
Review focus: RAG architecture, hybrid retrieval, language detection, evaluation.

Reviewed by Nishant Agarwal, Full Stack Developer. Built NovaMind’s APIs, real-time layer and deployment.
Review focus: Express APIs, WebSockets, authentication, security, deployment.

A customer starts a question in English, finishes it in their first language, then asks a follow-up by voice. Most chatbots get one of three things wrong: they lock onto the wrong language, ask the customer to pick one from a menu, or answer from general internet knowledge instead of the company’s actual policy.

NovaMind AI is a multilingual chatbot assistant built to get all three right. Users speak or type, switch languages mid-conversation, and get answers retrieved from the business’s own FAQs, SOPs and policy documents, with the source shown. NovaMind launched with English, Hindi and Hinglish (mixed Hindi-English), the languages its client’s users speak; the same pipeline is designed for adding more. This case study explains how our team built it, where retrieval-augmented generation (RAG) did the heavy lifting, what we measured, and what we would do differently.

TL;DR

A four-person THE TISA team built NovaMind AI in two months for US-based client Albert Joe. It is a multilingual voice and text assistant that detects each message’s language automatically, retrieves answers from each business’s own documents using hybrid keyword and vector search, cites its sources, and hands it off to a person when the evidence is weak. Over the measured period, NovaMind handled 2,000+ conversations; about 50% involved non-English or mixed-language messages, around 40% included voice, and uptime was approximately 99.5%.

It’s a working example of the RAG development services we deliver: retrieval designed for accuracy, built into a complete application rather than bolted onto a chatbot.

NovaMind AI at a glance

What problem did the client need to solve?

Albert Joe needed an assistant that follows how his users actually talk: in more than one language, often mixed in a single sentence, by voice as often as by text, and about the business’s own products and policies. Four gaps made off-the-shelf chatbots a poor fit.

This is a mainstream US problem, not a niche one. More than one in five US residents speak a language other than English at home, according to the U.S. Census Bureau’s American Community Survey. Many of them switch languages mid-sentence, whether that’s Spanish and English or, for this client’s users, Hindi and English. An assistant that assumes one language per message fails exactly those customers.

Why did NovaMind need RAG, not just a better chatbot?

Language fluency doesn’t tell a model what your refund window is. A large language model can only answer reliably about your business if it can see your business knowledge at the moment it answers. That’s the core idea behind retrieval-augmented generation: find the relevant passages first, then generate an answer from them.

For NovaMind, RAG gave us three things a standalone model couldn’t:

• Current answers. When a policy document changes, NovaMind re-indexes it and the answers change the same day. No retraining.
• Checkable answers. Every answer can cite the document it came from, so users and support staff can verify it.
• Separate knowledge per customer. Each business on the platform has its own knowledge base, and retrieval never crosses between them.

This is the same approach we take in every engagement under our custom RAG development services: start from the business knowledge, then design retrieval around it.

What did THE TISA build?

We built NovaMind as a complete AI application, then connected it to each business’s knowledge and systems. Buyers often budget only for the AI layer. In practice, the accounts, sessions, history, voice states and error handling around it took as much engineering as the retrieval itself.

How does a RAG assistant compare with other chatbot approaches?

A RAG assistant built into the application can find, cite and act on business knowledge. A generic chatbot can only talk about it. Here’s how the common options handled the jobs NovaMind needed to do.

How does the NovaMind RAG pipeline work?

NovaMind runs two pipelines. Ingestion prepares documents whenever an admin adds or changes one. Answering runs for every question, voice or text, and always filters to the asking customer’s own knowledge before it searches.

1. Preparing business documents

Document preparation is an important part of retrieval quality. FAQs, SOPs, product sheets, and policy documents are uploaded by each customer’s admin, split into chunks that retain their section headings, and tagged with tenant ID, document version, and language. A proposed parsing configuration uses PyMuPDF for PDFs and python-docx for Word documents, with 600-token chunks and a 100-token overlap. Admins can test questions against a document before it goes live to identify gaps in the source material.

2. Hybrid retrieval: exact terms plus meaning

Support questions mix exact identifiers with loose phrasing. “Cancel policy for Pro?” needs the exact plan name and the meaning of “cancellation policy”. So NovaMind searches both ways in Elasticsearch and merges the results.

3. Grounded answers with citations

The model receives the question, the relevant conversation turns and the retrieved passages, with instructions to answer only from those passages and to cite them. LlamaIndex handles indexing and citation mapping; LangChain coordinates prompts and memory. A simplified version of the retrieval guard looks like this:

# Illustrative, not production code
def retrieve(question, tenant_id, k=5):
    filters = {"tenant_id": tenant_id, "is_current": True}   # tenant first
    hits = hybrid_search(question, filters, limit=20)          # keyword + vector
    passages = rerank(hits)[:k]
    if not passages or passages[0].score < MIN_SCORE:
        return None        # caller clarifies, declines or hands off
    return passages

4. Knowing when not to answer

A wrong answer with a citation looks trustworthy, which makes it worse than no answer. When the best passage scores below the threshold, NovaMind asks a clarifying question, says it doesn’t have that information, or offers a person. It never falls back to general knowledge for policy questions.

What the model does, and what our code controls

Why we split the work this way

Letting the model search, decide and act looks impressive in a demo and is hard to trust in production. So the model handles the fuzzy part, understanding what a question means and phrasing a clear answer, while code handles anything with a right answer: which customer’s documents are searched, whether the evidence is strong enough, and what the user is allowed to do. The trade-off was more engineering up front, in exchange for testable rules and an audit trail for every answer.

What NovaMind never does on its own: answer a policy question from general knowledge, search another customer’s documents, change data without the user’s confirmation, or hide that it handed off to a person.

Developer’s perspective: making retrieval work for mixed-language questions

“Our retrieval approach needs to handle questions such as ‘Am I eligible for a refund if I cancel my Pro subscription?’ without losing the plan name or confusing cancellation terms with refund eligibility. Keyword search helps match exact identifiers, while semantic search handles paraphrased questions. When the retrieved evidence is insufficient, the system should say ‘I don’t know.’”

— Anuj Kumawat, AI/ML Engineer, THE TISA

How does NovaMind handle multiple and mixed languages?

The language-detection approach combines script, vocabulary and sentence-pattern signals to identify mixed-language messages before retrieval and generation.

Generic language detectors pick one language per message and assume the script tells you the language. Mixed-language messages break both assumptions: Hindi typed in the Latin alphabet, for example, looks like English to most tools. So the module combines three signals, shown here with examples from NovaMind’s launch languages:

Because detection runs per message, a user can start in English, switch languages, and switch back; the reply follows each time. Retrieval still searches the mostly English documents, and the answer is written in the language the user used. Adding a language means adding its detection signals, speech models, test questions and evaluation, not just translating the interface.

How do memory and voice stay fast?

Conversation memory. History is stored in MongoDB Atlas, but only the relevant recent turns are sent with each new question. Sending the full history slowed answers, raised cost and sometimes pulled replies off-topic. Redis caches session context for quick follow-ups.

Voice. Voice and text share one pipeline, so a spoken question gets the same retrieval, citations and safety checks as a typed one. Speech is transcribed, processed like text, streamed to the screen over WebSockets as it’s generated, then read aloud. The interface shows listening, processing and speaking states, and any failed step offers a retry instead of a silent pause.

Developer’s perspective: keeping voice fast enough to feel like a conversation

“Voice has six steps between the user finishing a sentence and hearing a reply. Run them strictly in sequence and the pause feels broken. We stream text over WebSockets as soon as the first tokens arrive and begin speech synthesis after the first complete sentence, bringing the first spoken word to about 2.5 seconds.”

— Nishant Agarwal, Full Stack Developer, THE TISA

How do tool calling and human handoff work?

RAG answers questions about documents. Questions about live data, such as “where is my order?”, need an API call. NovaMind can request one of a short list of approved actions; our backend, not the model, makes the call with the signed-in user’s permissions, and any action that changes data needs the user’s confirmation. This follows the patterns from our AI workflow automation work: the model chooses what to do, and code decides what’s allowed.

When evidence is weak, the request is out of scope, or the user asks for a person, NovaMind hands the conversation to a human with the full transcript and the passages it retrieved, so the customer never repeats themselves.

How does NovaMind protect private conversations and business knowledge?

In a multi-tenant RAG system, the most likely leak isn’t an attacker. It quietly retrieves a passage from another customer’s documents. NovaMind filters on tenant ID before any search is scored, so other customers’ chunks are never candidates.

• Password hashing, JWT sessions, protected APIs and conversation records
• Input validation, request sanitisation, rate limiting and CORS protection
• Model API keys held server-side only; HTTPS throughout; secure MongoDB Atlas connections
• Only the question, relevant turns and retrieved passages are sent to model providers, never whole documents
• Refusals, handoffs and tool calls logged for audit

NovaMind does not claim SOC 2, HIPAA or other certifications. Data residency, retention and deletion follow the client’s agreement; for provider-side settings, see OpenAI’s API data controls. Our AI SaaS security and compliance guide covers the questions US buyers should ask, and you can read how THE TISA protects data.

What US buyers should ask before deploying a RAG assistant

1. Where is retrieval filtered by the customer? It should happen in the search itself, before scoring, not only in the interface.
2. What leaves your system? Know exactly which fields reach the model provider, and whether names or account data can be left out.
3. What does the provider keep? Check retention and training settings for the specific API you use.
4. Can every answer be traced? You should be able to see the source document, its version and who approved any action.
5. Which privacy laws apply? CCPA and similar state laws depend on your users and data; your legal team should confirm.

How did we deliver NovaMind in two months?

We followed the same nine-step process we use on every RAG engagement. Two months was possible because the knowledge base was small and well organised; messier documents or more integrations would extend it.

What results has NovaMind delivered?

NovaMind handled 2,000+ conversations from June 2024 to August 2026. Around 50% involved non-English or mixed-language messages, while approximately 40% included at least one voice message.

How we evaluate answer quality

Request success is not answer accuracy. A request can return without an error and still be wrong, so we evaluate retrieval and answers separately, using LangSmith traces and a test set of real questions in all three languages, including questions the documents can’t answer

What changes for each role

These describe what the platform makes possible. The measured effect on support workload is the “Resolved without a human” row above.

What went wrong, and what we changed

• The transliterated text looked like English. Early replies came back in the wrong language because Hindi typed in the Latin alphabet was classified as English. We replaced generic detection with the script, vocabulary and pattern module above.
• Long histories hurt answers. Sending the full conversation slowed responses and drifted off-topic. We now send only relevant recent turns.
• Messy documents produced weak answers. No prompt fixed a vague or outdated source. The admin test-before-publish flow moved quality control to the document owner.
• We started measuring too late. Some baselines we wanted were never recorded. On new projects, evaluation and usage tracking go live before launch.

What the client said

“We wanted an AI platform where users could speak or type in English, Hindi, and Hinglish without selecting a language each time. Team TISA understood this requirement clearly and developed NovaMind AI around the way our users communicate. Users can switch languages, use voice, and continue conversations without interruption. We are impressed with the final platform and satisfied with how the team managed the development process.”

— Albert Joe, Founder, NovaMind AI

What should US teams learn from NovaMind?

The main lesson: retrieval quality and knowing when not to answer matter more than which model you choose. These are the six things we’d tell any team planning a support or knowledge assistant.

1. Start with your documents, not the model. Clean, current, well-structured sources improved answers more than any prompt change.
2. Use hybrid search. Customers type exact plan names and order numbers as often as loose descriptions.
3. Test in the languages customers actually use. An English-only test set hides your worst failures with customers who mix languages, whether Spanish-English or Hindi-English.
4. Isolate customers at retrieval time, not only in the interface.
5. Design the handoff before launch, so customers never repeat themselves.
6. Measure from day one, including a baseline of support volume and resolution time.

How to estimate the value of a RAG support assistant

Most of the return comes from support time saved on questions the documents already answer. You can estimate it in five minutes:

Annual value = Monthly conversations × Share resolved without a human × Minutes per human-handled conversation ÷ 60 × Loaded hourly cost × 12

Measure your own baseline before launch so you can show the real change afterwards. Our RAG development services start with exactly that baseline.

When is a RAG assistant like NovaMind the right fit?

RAG pays off when customers ask many questions your documents already answer, and those documents change too often to retrain a model. It’s the wrong starting point when:

• Answers aren’t written down. RAG can only retrieve what exists; document the knowledge first.
• Question volume is low. A clear FAQ page and a human inbox are cheaper.
• Most requests need live account data. That’s mainly an AI integration and agent problem, with retrieval as a smaller part.

Is your business ready for a RAG assistant?

• Your top 50 customer questions are answered somewhere in writing
• Someone owns keeping those documents current
• You know which data may and may not go to an AI provider
• A human team can receive handoffs
• You know which languages your customers actually write in
• You can measure today’s support volume and resolution time

Checked fewer than four? Start with a knowledge source audit, the second step of our RAG development company process, before choosing a model.

What’s next for NovaMind?

These are planned, not shipped: more Indian and international languages; WhatsApp, Slack and Microsoft Teams channels; CRM write-back with approval workflows; team accounts with role-based access; and analytics for retrieval quality and containment rate.

Frequently asked questions

Q1. Can a RAG chatbot support multiple languages, including customers who mix them?
Ans. Yes, but it takes more than a multilingual model. NovaMind detects the language of each message, retrieves from mostly English documents, and replies in the user’s language. Mixed and transliterated text needed custom detection because generic tools mistake it for English. NovaMind launched with English, Hindi and Hinglish; the same approach applies to other language pairs, such as Spanish and English, with its own test set.

Q2. How long does it take to build a RAG assistant like NovaMind?
Ans. NovaMind took two months with a team of four, including voice, three languages, tool calling and human handoff. A text-only assistant over one well-organised knowledge base can take less; messy documents, more languages or deep integrations add time.

Q3. How much does a multilingual RAG assistant cost?
Ans. Cost depends on document volume and quality, languages, voice, integrations and security requirements, plus ongoing model, speech and hosting costs. A scope like NovaMind’s typically falls in the USD 15k–50k range to build. See our AI SaaS development cost guide for the main drivers.

Q4. How is RAG different from fine-tuning?
Ans. RAG retrieves relevant passages from your documents when a question arrives; fine-tuning changes the model through extra training. For support and knowledge assistants, RAG is usually the better fit: update a document and answers change the same day, and every answer can cite its source.

Q5. How do you stop a RAG assistant from making things up?
Ans. Answer only from retrieved passages, require citations, set a minimum retrieval score below which the assistant clarifies or declines, and hand off to a person when confidence is low. Then test groundedness and refusals on a real question set before launch.

Q6. Does customer data go to OpenAI or other model providers?
Ans. Only what each request needs: the question, relevant recent turns and the retrieved passages. API keys stay on the server. Check each provider’s current retention and training settings for the API you use.

Q7. Can a RAG assistant connect to our helpdesk or CRM?
Ans. Yes. NovaMind calls approved business APIs for live data and asks the user to confirm any change. Helpdesk and CRM connections follow the same pattern, each with its own permissions and tests.

Q8. Why does NovaMind use more than one model provider?
Ans. OpenAI handles most generations, with Claude available as an alternate path for long-context answers. Keeping providers behind one internal service lets us switch as prices and quality change; we explain why in our vendor-agnostic AI guide.

Planning a multilingual RAG assistant?

If your customers ask the same questions every day and the answers already exist in your documents, a RAG assistant can take that load off your team, in the languages your customers actually use. We start the way we did with NovaMind: a short review of your documents, languages, data rules and the systems the assistant needs to reach.

Explore our RAG development services or book a free RAG strategy consultation. We reply within one business day.Related: Clove AI project management case study · OpsPilot AI integration case study · All AI case studies · AI for SaaS and technology companies