Irus AI started as a simple idea: what if you had a personal AI that knew everything about your work, could search across all your documents, and remember conversations over time?
The Architecture
We built Irus AI using a modern stack: - Frontend: Next.js 16 with App Router - AI Engine: LangChain for orchestration - Vector Database: Pinecone for semantic search - LLM: OpenAI GPT-4 for reasoning - Memory: Redis for conversation history
Key Features
Live Search Unlike traditional keyword search, Irus AI uses vector embeddings to understand the meaning behind your queries. Ask "what did I decide about the database last week?" and it finds the relevant conversation even if you didn't use those exact words.
Document Intelligence Upload PDFs, docs, or paste text — Irus AI chunks, embeds, and indexes everything. It can answer questions about your documents with citations.
Long-Term Memory Most AI chatbots forget everything after the conversation ends. Irus AI maintains a persistent memory graph, connecting related concepts across conversations.
Lessons Learned
- 01.Chunking strategy matters — We experimented with fixed-size, semantic, and recursive chunking. Recursive won for most use cases.
- 02.Embedding model choice — OpenAI's text-embedding-3-small gave us the best price/performance ratio.
- 03.Prompt engineering — The system prompt is 80% of the battle. We iterated dozens of times.
What's Next
We're adding multi-modal support (images, audio), agent capabilities (autonomous task execution), and a plugin system for custom integrations.
Try Irus AI at irus-ai.onrender.com