h2oGPT tutorial: chat with PDFs and web pages

Learn how the 2023 h2oGPT demo uses uploaded PDFs, URLs and pasted text for document questions, summaries and source references.

Player not loading? Watch on YouTube

Uploaded on July 10, 2023, this introduction by Andreea Turcu explains GPT concepts before demonstrating h2oGPT's chat and document workflows. No release number is supplied; Falcon 40b is the selected model near the end. The interface and capabilities described belong to this 2023 demo, and Turcu warns that the platforms will change.

The walkthrough starts with everyday prompts, then moves to the data source tab. Turcu uploads the team's roughly 22-page paper, selects My Data and requests a summary. She checks the source reference at the end of the answer to see whether it used the document. The initial response is generic, which leads to practical advice: ask specific questions and start a fresh conversation when changing topics.

A second example adds a Wikipedia URL about Norwegian Forest cats and asks for a summary and follow-up details. Pasted text provides another input route. The demo also covers chat export and import, model selection and expert controls such as early stopping and penalties. A web search toggle is described as experimental but is not tested.

For viewers considering self-hosted local AI, Turcu describes deployment on premises or in private clouds, with a chatbot interface and Python API. She presents control over data and customization as benefits of open source deployment. These are the speaker's claims; the video demonstrates hosted interfaces rather than an installation procedure or a privacy audit.