How this works

How it works

1. Search. The papers are split into short passages. Your question is compared with all of them to find the closest ones (retrieval).
2. Three AI agents. A Researcher checks whether the passages are enough and can search again with better words. A Writer answers using only those passages, with a [number] after each claim. A Checker verifies each claim against its source and sends the answer back if something is not supported.
3. Guardrails. Off-topic questions, attempts to change the rules ("prompt injection") and requests for personal data or advice are refused. If a claim cannot be verified, no answer is given.

Quality was measured on a fixed set of 40 test questions plus 16 "trap" questions, with the open-source evaluation tool Ragas. The full method, scores and code are in the project's documentation.

Built with

Python, LangChain and LangGraph, Google Gemini (answers), a small open embedding model (search), Chroma (vector store), FastAPI, and plain HTML/CSS/JavaScript.

Privacy

• No accounts and no cookies. The page font is loaded from Google Fonts.
• Your question, and short passages from the papers, are sent to Google's Gemini API (free tier) to write the answer. Google may use this text to improve its products, so please don't type personal information.
• For quality monitoring, questions and answers are recorded in LangSmith and kept for 14 days.
• To keep the service free, each visitor can ask a limited number of questions per day.

About me

I'm Alzayat Saleh, a deep learning engineer and Research Fellow at AIMS@JCU (James Cook University). I build computer vision for marine science, aquaculture and agriculture.

alzayat.com · Google Scholar · GitHub

The papers

Loading the list…