All Projects

Portfolio AI – RAG-Powered Personal Portfolio

This portfolio site, featuring an embedded RAG-powered chatbot that answers questions about my background, skills, and projects. Custom ingestion pipeline, pgvector retrieval, GPT-4.1-mini generation with anti-hallucination guardrails, and SSE streaming.

Full-Stack Engineer2024 – Present
GitHub Live Demo

Stack

Next.js 14TypeScriptTailwindCSSFastAPIpgvectorPostgreSQLOpenAIPython

Portfolio AI – RAG-Powered Personal Portfolio

Overview

My personal portfolio site features an embedded AI chatbot ("Anvar AI") that answers questions about my background, skills, and projects using retrieval-augmented generation. All answers are grounded in my actual resume and curated project data — hallucinations are significantly reduced through RAG, structured context retrieval, and constrained prompts.

Architecture

Frontend (Next.js 14)

  • App Router with TypeScript and TailwindCSS
  • Pages: Home, Projects, Experience, Contact, Resume
  • Floating chat widget on every page with streaming responses
  • Citations displayed under each AI response

Backend (FastAPI + OpenAI)

  • FastAPI server with streaming SSE chat endpoint
  • Custom RAG pipeline: chunking, embedding, cosine similarity retrieval
  • pgvector (PostgreSQL) as the vector store
  • OpenAI-compatible API for embeddings and chat completion — backend can be pointed at Azure OpenAI, Ollama, or any compatible provider without code changes

RAG Pipeline

  • Ingestion: resume, project markdown files, and about content are chunked (800 chars, 120 overlap), embedded via text-embedding-3-small, and upserted into pgvector with deterministic IDs for idempotent re-runs
  • Retrieval: cosine similarity search with IVFFlat ANN index, top-5 chunks returned with metadata
  • Generation: GPT-4.1-mini with a strict system prompt containing 8 anti-hallucination rules, 6-turn conversation history, and numbered context blocks
  • Streaming: SSE events deliver citations first, then tokens one at a time for a typewriter effect

Key Highlights

  • Anti-hallucination guardrails: model only uses retrieved context, never invents employers or projects
  • Citations in every response with source metadata
  • Streaming typewriter effect for AI responses
  • Configurable LLM backend (OpenAI, Azure, Ollama) via environment variable
  • One-command setup with docker-compose