Upload your business data. Ask anything in plain language. Get instant AI-powered insights.
NodeTalk is a production-grade multi-tenant SaaS platform that enables non-technical teams to interact with their CSV data using natural language — powered by a custom RAG (Retrieval Augmented Generation) pipeline.
As a Marketing Manager:
"I uploaded last quarter's sales CSV and asked 'Which product had the highest revenue in February?' — NodeTalk answered instantly with exact numbers. No SQL, no waiting for analysts."
As an Operations Head:
"I asked 'Which city had the most delayed deliveries?' — the AI analyzed my logistics data and gave me a ranked list in seconds."
As a Team Owner:
"I invited my team members, assigned roles, and now everyone can query our shared data — securely, without exposing raw files."
Persistent AI workspace for teams — large-file processing, vector search, multi-tenant collaboration, RBAC, persistent conversations, and API integration.
Built for data control and scalable AI workflows, rather than one-off conversational sessions.
Backend: Node.js, TypeScript, Express.js, PostgreSQL + pgvector, MongoDB, Redis, AWS S3, JWT, Zod
AI / RAG: Hugging Face Embeddings, all-MiniLM-L6-v2, pgvector, Groq LLM (gpt-oss-20b), Custom RAG Pipeline
Mobile: Flutter, Dart, Riverpod, Dio, SharedPreferences, Material 3
Cloud & DevOps: Docker, AWS ECR, ECS Fargate, AWS RDS, AWS ALB, Upstash Redis, GitHub Actions
Step 1: CSV Upload
User uploads CSV → Stored in S3/MinIO
DB record created (datasource)
Step 2: Embedding (Indexing)
Each CSV row converted to text:
"product_name: iPhone 15, sales: 150, month: January"
↓
HuggingFace API → 384-dimensional vector
↓
Stored in PostgreSQL pgvector table
Step 3: Query (Retrieval)
User asks: "Best selling product?"
↓
Query → HuggingFace → query vector
↓
pgvector cosine similarity search
Top 5 most relevant rows retrieved
Step 4: Generation
Retrieved context + conversation history
↓
Groq LLM (openai/gpt-oss-20b)
↓
Natural language answer ✅
Step 5: Persistence
User message + AI answer saved to DB
Conversation history maintained
Flutter App (iOS/Android)
↓
AWS ALB (Load Balancer)
↓
AWS ECS Fargate (Node.js API)
↓ ↓ ↓ ↓
AWS RDS MongoDB Upstash AWS S3
PostgreSQL Atlas Redis (Files)
+ pgvector (Chats) (Sessions)
↓
HuggingFace API Groq API
(Embeddings) (LLM Chat)
Every organization's data is completely isolated:
→ org_id on every table
→ JWT contains orgId — injected by middleware
→ Every DB query filters by org_id
→ No cross-tenant data leakage
Role-Based Access Control:
OWNER → Full access, invite members, delete org
ADMIN → Manage members, upload data
MEMBER → View and query data only
JWT Security:
→ Access token: 15 minutes
→ Refresh token: 7 days
→ Logout: refresh token blacklisted in Redis
→ Token rotation on every refresh
# Clone
git clone https://github.com/sumit2393/NodeTalk.git
cd NodeTalk
# Start Docker services
docker compose -f infra/docker-compose.yml up -d
# Install dependencies
cd apps/api && npm install
# Environment
cp .env.example .env
# Fill in your keys
# Run
npm run devServices started by Docker:
- PostgreSQL + pgvector → localhost:5432
- MongoDB → localhost:27017
- Redis → localhost:6379
- MinIO (S3) → localhost:9000
NODE_ENV=development
PORT=3000
PostgreSQL= your secret keys
MONGO_URI=localhost
Redis=your key
JWT=your secret
AWS / MinIO= your keys
# AI
GROQ_API_KEY=your_groq_key
HUGGINGFACE_API_KEY=your_hf_key✅ Multi-tenant architecture — org-level data isolation
✅ JWT with Redis blacklisting — secure token management
✅ Custom RAG pipeline — not just API integration
✅ pgvector — no separate vector DB needed
✅ Dockerized — same environment everywhere
✅ AWS ECS Fargate — serverless, auto-scaling
✅ Role-based access control — enterprise ready
✅ Error handling — graceful failures
✅ Environment validation — Zod schema on startup
✅ Conversation persistence — full chat history
MIT © 2026 Sumit Kumar
Built to showcase production AI engineering skills — RAG pipeline, multi-tenant SaaS, cloud deployment, and mobile integration.



