Inside the portfolio
From question to answer.
A RAG flow with a model that can request four limited tools. Contact actions are never performed autonomously: the visitor must submit the form.
The model chooses whether to answer or use a tool. Select nodes for details; on mobile you can scroll and zoom the graph.
LLM model
OpenAI generates text or selects a tool based on intent. Vercel AI SDK orchestrates at most two steps, with store: false and streamed generation.
A short cycle, not an unlimited agent
The model receives descriptions of the four tools, can request their execution and receives the result. Orchestration stops within two steps. Informational questions normally receive text only.
Two paths, two treatments
AI calls receive question and context. The contact form uses a separate endpoint: its fields are never inserted into the prompt or history.
Limits and visitor control
You can stop streaming, open or download the resume and choose when to enter the exhibition or send a contact. RAG improves grounding in documents, but does not make answers or tool selection infallible.