Multilingual RAG and Conversational AI

NovaMind AI

Multilingual RAG Voice and Text Assistant

NovaMind AI supports voice and text conversations in English, Hindi, and Hinglish. It detects each message’s language, retrieves answers from business documents using hybrid keyword and vector search, cites its sources, and hands off to a person when the evidence is weak.

USAMarket

5 Months (February–June 2024)Timeline

4 MembersProject Team

Artificial IntelligenceIndustry

Executive Summary

NovaMind AI

Client: Albert Joe
Market: USA
Industry: Artificial Intelligence
Timeline: 5 Months (February–June 2024)
Team: Project Team - 4 Members
Services: AI Agent Development, Generative AI Development, RAG Development Services, AI Application Development
Technology: Next.js, Express.js, REST APIs, MongoDB Atlas
Primary Outcome: Around 85% of registered users Monthly Active User Rate

Verified project metrics

Project Outcomes

From June 2024 through August 2026, NovaMind recorded 2,000+ conversations. Approximately 50% were non-English or mixed-language, 40% used voice, 55% of users returned in two or more separate weeks, and the monthly active-user rate was approximately 85%. These usage and reliability figures are distinct from answer accuracy. Retrieval hit rate, groundedness, correct refusals, language match, and resolution without a human are being benchmarked.

Around 85% of registered users

Monthly Active User Rate

2,000+

Total Conversations

Nearly 40% voice and 60% text

Voice and Text Usage

Around 50% of conversations

Non-English and Mixed-Language Share

Nearly 55%

Returning Users

Nearly 92%

Voice-to-Text Success Rate

Project Overview

NovaMind AI is a multilingual voice and text assistant built for customer support and internal knowledge. Users can switch between English, Hindi, and Hinglish during the same conversation.

THE TISA developed the platform for Albert Joe, Founder of NovaMind AI, with a four-person team. The discovery-to-launch timeline ran from February to June 2024.

The application connects per-tenant business knowledge, hybrid retrieval in Elasticsearch, source citations, approved API actions, and human handoff. Conversation history and voice processing are part of the same application.

From June 2024 through August 2026, the platform recorded 2,000+ conversations, approximately 50% non-English or mixed-language use, approximately 40% voice use, and approximately 99.5% uptime. Answer-quality benchmarks are in progress.

What We Built

We built a complete multilingual RAG application with Next.js voice and text controls, authentication, sessions, conversation history, streamed replies, and responsive layouts. The knowledge layer includes per-tenant document collections, hybrid keyword and vector retrieval in Elasticsearch, source citations, re-indexing when documents change, and an admin upload and test flow. LangChain coordinates prompts and memory, LlamaIndex supports indexing and citation mapping, LangGraph coordinates approved API actions and human handoff, and LangSmith supports tracing and evaluation. OpenAI GPT-4o handles generation, with Claude available as an alternate path for long-context answers.

Client Background

Albert Joe, Founder of NovaMind AI, needed an assistant that could support users who switch languages, mix languages within a message, and alternate between voice and text. The assistant needed to answer from the business’s actual policies and procedures, retain relevant context for follow-up questions, and pass unresolved requests to a person with the conversation context.

Project Scope

We worked on the complete project, from planning and design to deployment, including:

Requirement gathering
Conversation and user-flow planning
Voice, text, and multilingual interaction planning
UI/UX design
Frontend development
Backend and REST API development
User authentication
MongoDB Atlas setup
OpenAI, RAG, and LangChain implementation
Python-based AI processing
Conversation memory
Speech-to-text and text-to-speech setup
Real-time response streaming
Performance optimisation
Security implementation
Cross-browser and device testing
VPS configuration, deployment, and live review

Target Audience

NovaMind AI is built for businesses that want to use AI for customer support or daily internal work. The main users include:

Product teams
Customer support teams
Operations teams
Service-based businesses
SaaS companies
Enterprises
Businesses serving English, Hindi, and Hinglish-speaking users

Objectives

The main objectives were to:

01Create an AI platform for voice and text conversations
02Support conversation in English, Hindi, and Hinglish
03Identify the user’s language automatically
04Allow users to switch languages during a conversation
05Give answers based on previous messages
06Save and manage conversation history
07Find useful information from connected business data
08Provide fast, real-time responses
09Avoid repeated or unrelated answers
10Protect user accounts and conversation data
11Support future integrations and new AI features

Business Challenge

Before development, we looked at the main problems users face when talking to AI in different languages.

01Basic chatbots often struggle when users switch between languages. Sometimes, users have to select a language manually or repeat their message.
02Voice conversations were also more complex. The system had to convert speech into text, understand the language and question, and then give the reply in voice.
03The platform also needed to remember earlier messages so users could ask follow-up questions without repeating the same details.
04The main challenge was to bring language detection, voice support, conversation memory, and useful AI responses into one simple platform.

Our Solution

We developed a multilingual AI platform for voice and text conversations.

Detect the language of each message, including mixed-language and Roman-script Hindi input.
Retrieve from the asking customer’s own knowledge base using hybrid keyword and vector search in Elasticsearch.
Rank the search results and provide the top five passages as evidence for answer generation.
Generate answers from retrieved passages and map citations back to their source documents.
Re-index updated business documents and allow administrators to test questions before publishing knowledge changes.
Store conversation history in MongoDB Atlas and send only relevant recent turns with each new question; Redis caches session context.
Stream text over WebSockets and provide speech input and spoken replies within the same conversation workflow.
Run approved business API calls through the backend, enforce permissions, and request user confirmation before changing data.
Hand off unsupported or out-of-scope requests to a person with the transcript and retrieved passages.

Before → After

How Conversations Changed

Before

Users mainly depended on text-based chatbots

After

Users can communicate through voice or text

Before

Language had to be selected manually

After

The system detects English, Hindi, Hinglish, and Roman Hindi automatically

Before

Language changes could interrupt the conversation

After

Users can switch languages during the same conversation

Before

Follow-up questions often lost earlier details

After

Previous messages help continue the conversation

Before

Answers mainly depended on general AI knowledge

After

The system uses connected business information when needed

Before

Users had to wait for the complete response

After

Answers appear on the screen in real time

Before

Voice processing required separate tools

After

Voice input, AI response, and playback work in one flow

Before

Earlier discussions were difficult to continue

After

Conversation history is saved across sessions

Delivery ownership

THE TISA's Role

Product Strategy

Discovery

UX/UI Design

AI Architecture

Backend Engineering

Frontend Engineering

Mobile Development

Integrations

QA

Key Features

The platform includes the following features:

03Voice and Text Conversations

Voice and Text Conversations

Chat through voice or text
Switch between speaking and typing
Receive replies in text or voice

Development Process

Mapped user languages, voice and text interactions, and questions requiring documents versus live business data. Audited FAQs, SOPs, and policies for gaps and outdated versions.

Technical Challenges

Transliterated Hindi was classified as English

Replaced generic detection with combined script, vocabulary, and sentence-pattern signals.

Long conversation histories slowed requests and introduced unrelated context

Send only relevant recent turns and cache session context in Redis.

Vague or outdated documents produced weak answers

Added an administrator test-before-publish workflow and re-indexing when source material changes.

Some business baselines were not recorded before launch

Separate platform-use metrics from answer-quality evaluation and keep the quality benchmark marked as in progress.

Technology Stack

Next.js
Express.js
REST APIs
MongoDB Atlas
OpenAI GPT-4o
Claude
Hugging Face
LangChain
LangGraph
LlamaIndex
LangSmith
Python
JWT Authentication
WebSockets
Redis
Elasticsearch
PostgreSQL
Ubuntu VPS
Nginx
PM2
SSL
GitHub

Design Decisions

The platform was designed to keep voice and text conversations simple and easy to follow. Important decisions included:

Conversation-focused interface
Voice and text controls in one place
Clear user and AI messages
Easy access to previous conversations
Visible listening and processing states
Automatic language detection
Real-time response display
Simple login and account flow
Responsive layouts across devices
Clear error and retry messages
Only essential controls in the chat area
Easy voice playback for generated responses

SEO and Indexing

NovaMind AI includes a public website and private user conversation pages. SEO work focused on the public website, while private account and conversation pages were kept out of search results. The work included:

Clear page titles
Proper page headings
Relevant AI platform keywords
Multilingual AI-related content
Voice assistant-related content
Clean page URLs
Meta descriptions
Mobile-friendly layouts
Image optimisation
Fast public-page loading
Restricted indexing for private pages
Clear feature and technology sections

Performance Optimisation

The implementation coordinates conversation context, response streaming, voice processing, and error recovery. Non-AI endpoint response time is reported separately from answer quality and complete voice-response latency.

Stream responses over WebSockets as tokens arrive
Send relevant recent conversation turns instead of the complete history
Cache session context in Redis
Show listening, processing, and speaking states
Provide clear error messages and retry options
Use responsive layouts across devices
Trace retrieval and AI requests in LangSmith
Operate application services on an Ubuntu VPS with Nginx and PM2

Security

Tenant filtering runs inside retrieval before scoring. Application controls protect accounts and private conversations, while backend code enforces permissions and confirms actions that change data. No SOC 2, HIPAA, or other certification is claimed for NovaMind.

Tenant filtering before keyword or vector search scoring
Password hashing and JWT sessions
Protected APIs and conversation records
Input validation, request sanitization, and rate limiting
CORS configuration for browser requests
Server-side model API keys
HTTPS and secure MongoDB Atlas connections
Only the question, relevant turns, and retrieved passages sent to model providers
User confirmation before actions that change business records
Refusals, handoffs, and tool calls logged for audit
Data residency, retention, and deletion governed by the client agreement

Architecture Diagram

Results

For Users

01Users can speak or type in English, Hindi, and Hinglish, switch language during a conversation, and ask follow-up questions with relevant context. Answers include business-source citations, and unresolved requests can be handed to a person with the transcript.

For the Client

01The client received a multilingual RAG application connecting customer-specific knowledge, approved API actions, voice and text interactions, and human handoff. The four-person project ran from February to June 2024.

Project Showcase

Visual Identity

NNovaMind AI

Inter

NovaMind AI

Color

#7C3AED
01Primary Purple124 58 237
#8B5CF6
02Bright Violet139 92 246
#050816
03Dark Background5 8 22
#0F172A
04Secondary Background15 23 42
#111827
05Card Background17 24 39
#FFFFFF
06White255 255 255
#E2E8F0
07Light Text226 232 240
#94A3B8
08Secondary Text148 163 184
#3B82F6
09Blue Accent59 130 246

Type

01 · Primary4 weights
Inter
ABCDEFGHIJKLMNOPQRSTUVWXYZ
abcdefghijklmnopqrstuvwxyz
0123456789
- Regular
- Medium
- Semibold
- Bold

Client Testimonial

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 JoeFounder, NovaMind AI

Future Scalability

Planned capabilities, distinct from the current implementation:

01
Additional Indian and international languages
02
WhatsApp, Slack, and Microsoft Teams channels
03
CRM write-back with approval workflows
04
Team accounts with role-based access
05
Analytics dashboards for retrieval quality and containment rate

Lessons Learned

While building NovaMind AI, we learned that:

01
Start with clean, current, well-structured business documents
02
Use hybrid search for both exact identifiers and paraphrases
03
Test in the languages and mixed-language patterns customers actually use
04
Apply tenant isolation during retrieval, not only in the interface
05
Design human handoff with the transcript and retrieved passages
06
Send relevant recent turns rather than the entire conversation history
07
Record a business baseline and start evaluation before launch
08
Keep API request success separate from factual answer accuracy

Conclusion

NovaMind AI connects multilingual voice and text conversations to business knowledge through tenant-scoped hybrid retrieval, source citations, contextual follow-ups, approved API actions, and human handoff. For a similar assistant, begin with a review of documents, languages, data rules, required systems, and the business outcome to measure. Explore THE TISA’s RAG development services or discuss your project through https://www.thetisa.com/get-in-touch.

Building Something Similar?

Talk with THE TISA about the architecture, timeline, team, and technology required for your project.

Discuss Your Project Explore More Case Studies

NovaMind AI

Multilingual RAG Voice and Text Assistant

NovaMind AI

Project Outcomes

Monthly Active User Rate

Total Conversations

Voice and Text Usage

Non-English and Mixed-Language Share

Returning Users

Voice-to-Text Success Rate

Project Overview

What We Built

Client Background

Project Scope

Target Audience

Objectives

Business Challenge

Our Solution

Before → After

THE TISA's Role

Product Strategy

Discovery

UX/UI Design

AI Architecture

Backend Engineering

Frontend Engineering

Mobile Development

Integrations

QA

Key Features

Voice and Text Conversations

Development Process

Technical Challenges

Transliterated Hindi was classified as English

Long conversation histories slowed requests and introduced unrelated context

Vague or outdated documents produced weak answers

Some business baselines were not recorded before launch

Technology Stack

Why These Technologies?

Design Decisions

SEO and Indexing

Performance Optimisation

Security

Architecture Diagram

Results

For Users

For the Client

Project Showcase

Visual Identity

Color

Type

Client Testimonial

Future Scalability

Lessons Learned

Conclusion

Related Services

Explore Similar Work

OpsPilot - AI-Powered Admin Portal

PeopleCore

Campus Flow

Building Something Similar?