Real-Time Research Assistant
What is this project?
A Perplexity-like AI research assistant built with CrewAI and the AG-UI protocol. It performs real-time web research with live agent event streaming, grounding responses with cited sources.
Why did I build it?
To explore real-time agent architectures and event-driven UIs while combining search, reasoning, and citation UX into a clean Next.js frontend.
How does it work?
- Uses the AG-UI protocol to stream backend agent events to the frontend UI in real time
- Research flow orchestrated by CrewAI
- Web search via EXA AI, intent classification with an LLM, and grounded answers with citations
Tech Stack
- AG-UI protocol
- CrewAI
- EXA AI
- Gemini 2.5 Flash (via OpenRouter)
- Next.js + Tailwind CSS
What were the biggest challenges?
- Designing a clear event schema for streaming multi-step agent updates
- Preventing UI flicker while preserving transparency of agent reasoning
- De-duplicating sources and normalizing URLs for cleaner citation cards
- Reliable intent detection between chat vs. search vs. exit
How can you use it or build on this?
- Use it as a research companion for quick topic deep-dives with sources
- Extend toolsets (scrapers, specialized APIs) and register them in the crew
- Adapt the UI to highlight confidence and provenance for enterprise use cases
Links
- Live Demo: https://ag-ui-crewai-researcher.folch.ai/
- GitHub: https://github.com/Folken2/ag-ui-crewai-research
What's next?
- Add deployment guide
- Expand crews (analysis/writing) and configurable tool registry
In the age of AI, stay curious!