Hiraya - AI-Powered Campus Assistant
Smart campus companion with RAG architecture
System Design & Network Topology
“Hiraya is an AI Campus Assistant for NIT Hamirpur that eliminates hallucinations using Retrieval-Augmented Generation. By compiling Hugging Face transformers locally in Node.js with PostgreSQL pgvector, I achieved a zero-cost embedding architecture and overcame Vercel serverless bundle limits through custom Webpack binary pruning.”
University students frequently struggle to find accurate, up-to-date information regarding campus rules, placements, and exam notices scattered across PDFs. Hiraya is an AI Campus Assistant for NIT Hamirpur using strict Retrieval-Augmented Generation (RAG) to provide verified, hallucination-free streamed answers from a maintainer knowledge base.
Full-stack Next.js (React 19) architecture. Frontend consumes server-sent token streams from LangChain and Groq LLM (Llama 3.1). The backend generates 384-dimensional text embeddings in-process using local Hugging Face ONNX models, stores vectors in PostgreSQL via pgvector and Prisma ORM, and executes fast cosine similarity retrieval.
Core Engineering Responsibilities & Contributions
Zero-Cost RAG Pipeline
Engineered a local in-process embedding pipeline using @huggingface/transformers (all-MiniLM-L6-v2), eliminating external embedding API costs completely.
Vector Database Integration
Integrated pgvector with PostgreSQL and Prisma to execute real-time cosine distance similarity queries on chunked campus data.
Streaming UI/UX
Built a responsive chat interface that decodes and renders text streams token-by-token directly from Groq and LangChain.
Secure Ingestion System
Developed a protected API route allowing maintainers to chunk and inject raw notices and PDF text directly into the knowledge base.
Key Architectural Decisions & Trade-Offs (Why X over Y)
Why Local Embeddings (transformers.js)?
External APIs like OpenAI introduce recurring costs and network round trips. Running all-MiniLM-L6-v2 locally in WebAssembly/ONNX reduced embedding costs to $0 and cut retrieval latency.
Why pgvector over Pinecone/Milvus?
For a campus system, a specialized vector DB adds unnecessary operational complexity. PostgreSQL with pgvector allowed relational and vector data to live together in a single Prisma schema.
Why Groq + Llama 3.1?
Groq LPU hardware generates tokens at extraordinary speeds. Combined with Next.js Edge streaming, it provides instantaneous response times.
Production Challenges & How They Were Solved
Serverless 250MB Bundle Limit on Vercel
The onnxruntime-node library ships with ~513MB of multi-OS binaries. Added Webpack outputFileTracingExcludes in next.config.ts to aggressively strip CUDA, macOS, and Windows binaries down to ~130MB.
Serverless Cold Starts on Read-Only Filesystems
Vercel serverless filesystems are read-only. Configured the 90MB Hugging Face model cache to download into /tmp/ and maintained an in-process singleton for warm execution.
Scalability & 100k Concurrent Users Roadmap
- 01.Migrate from serverless to persistent Docker containers (AWS ECS/Railway) to keep the 90MB embedding model permanently warm in memory.
- 02.Integrate Upstash Redis caching for frequently asked student questions (e.g., library timings, fee deadlines).
- 03.Incorporate multi-modal document parsing for scanned campus circulars.
Key System Highlights
- 80% response accuracy with NLP
- 90% faster information retrieval
- RAG architecture implementation
- 100% context-aware responses