Build a Video-Transcript Chatbot with MindStudio

Today I want to create an agent that lets you chat with one of my videos. OpenAI's custom GPTs have just arrived, but I want to show a different no-code service, MindStudio by YouAI, which is also the sponsor of this video. YouAI says it offers a free starting tier and paid features; the options shown here reflect the service in November 2023.

The platform calls these agents AIs. It lets creators publish them through a marketplace or elsewhere and potentially charge for access. I start a blank AI for one video, although the interface also offers templates for roles, content generation, talking to PDFs, data analysis and support.

On the prompt tab I describe the assistant as an instance of my Private GPT video and ask it to speak in my first-person style. Its job is to answer questions, explore the transcript and help apply what the video teaches.

I tell it to use only the transcript, cite relevant text and say 'I don't know' if the answer cannot be found. I also describe my tone as friendly, informative and casual. These instructions express the desired behavior; a prompt alone does not guarantee the model will always follow it.

The editor includes data sources for external knowledge, function calling, variables, warnings, a terminal and a debugger. Prompt templates can also incorporate dynamic information. I am not using function calls in this demonstration.

To create the knowledge source, I open the transcript on my Private GPT video, copy it into a text file and upload that file to a new data source. The interface creates embeddings and reports 13,824 vectors. It also offers views of the extracted text, chunks and vectors, along with a query tester.

Next I enable an automation. The first node collects the user's question and saves it as an original-query variable. A send-message step expands that question into three potential vector-search queries and saves the result as a refined-query variable; double curly braces let a prompt reference variables.

I add a query-data step that searches the video-transcription source using the refined query and stores the result. A final send-message step uses that result as context for the original question. It asks for relevant citations, forbids outside knowledge and says to admit when the supplied context does not contain the answer.

In message rules I select the transcript as a source and increase the maximum results to ten. For the underlying model I choose Claude V2, reduce temperature to 0.5 and set the maximum response size to 4,000. I also point out other models available in the 2023 interface, some of which required a paid plan.

I name the agent PrivateGPT 2.0 Chatter and add searchable metadata. Sharing settings let me make it public, allow or disable remixing and optionally protect access with a password. The editor also provides a simple landing page and analytics.

I test the draft by asking what the video is about, then ask for the installation commands. The assistant returns a summary and formatted commands, which look good to me. These examples show the workflow in action, though they do not establish that every retrieved answer or command will be correct.

The same design could use transcripts from all my videos rather than one. I also suggest a possible function-call workflow that fetches a transcript from a user-provided URL. Those are ideas for extensions, not steps implemented in this demonstration.