This video shows how to build a retrieval augmented generation chatbot on your own video data. Transform your in-house employee training videos into a chatbot that summarises the video & provides the correct video link. After this video, you can create chatbot with your own knowledge base.

If you are interested in chatting about your unique ideas to automate, drop a comment.

Apify Tool: [ Ссылка ]

If you want to automate your business processes, reach me out at: sharmabhuvnesh0610@gmail.com

Claim a free discover call here: [ Ссылка ]

Tools I use:
Botpress: [ Ссылка ]
Make: [ Ссылка ]
OpenAI: [ Ссылка ]
Perplexity: [ Ссылка ]
LinkedHelper: [ Ссылка ]

Cheers!

About the video:
This video tutorial demonstrates building a chatbot using retrieval augmented generation (RAG). The chatbot accesses a knowledge base of Loom video transcripts to answer user queries, providing summaries and video links. The tutorial details a three-step process: designing conversation flow, preparing data (including using Appify for automated transcript extraction), and integrating the data into the chatbot's knowledge base. The final steps show how to build the chatbot's conversational flow within a platform like Botpress.

Questions answered in the video:-
Q. What is the purpose of the chatbot described in this video?
A. The chatbot is designed to answer questions related to content found in a database of videos, such as Loom videos or internal training videos. It aims to provide users with quick answers extracted from video transcripts, along with a link to the specific video where the information is located. This is helpful when users don't want to watch an entire video but need specific information it contains.
Q. How does the chatbot find answers?
A. The chatbot utilizes a technology called Retrieval Augmented Generation (RAG). It has access to a knowledge base comprised of transcripts from videos. When a user asks a question, the system searches through the transcripts, identifies the relevant information, generates a summary and provides a link to the corresponding video. The search is semantic, not just keyword-based, meaning it understands the meaning of the query, rather than just matching specific words.

Q. What are the three main steps involved in building this chatbot?
A. The three main steps are:
Logical Flow: Determining the conversational flow, such as how the chatbot will greet the user and handle questions.
Data Preparation: Gathering all video links and creating a list.
Transcript Acquisition: Obtaining transcripts of all videos and formatting them into a suitable format for a knowledge base.

Q. How are video transcripts obtained, especially for a large number of videos?
A. For a small number of videos, transcripts can be manually downloaded from platforms like Loom. However, for large sets of videos, a tool called "Appify" is recommended. Appify offers a "Loom Transcript Grabber" tool that automates the process of extracting transcripts from video URLs and provides both the transcript and the video URL in JSON format. This is a bulk process and is very efficient.

Q. Why is the transcript data converted to JSON format?
A. JSON format is preferred because it structures the data in a way that is easy for Large Language Models (LLMs) to read. Specifically, the transcript is attached to the corresponding video URL, which allows the chatbot to both find the relevant information and link it back to the original video when answering user questions.

Q. How is the knowledge base built and structured?
A. The knowledge base is built within Botpress (or similar) and is structured to divide the transcript information into "chunks." This helps the system perform a more efficient search using semantic meaning, rather than simple keyword matching. Each chunk is linked back to its video source. This way, when a user asks a question, the chatbot identifies the most relevant chunk of text and can return the summary along with the associated video link.

Q.What agents and models are recommended and how can I change them?
A. It is recommended to use a Summary Agent and a Knowledge Agent for the system to work properly. Within the knowledge agent, it is best to select the "advanced model" and the "best model" in the question extractor model settings. In the summary agent, using the GPT-4 model (or equivalent) is recommended to get a good summary of the transcript data. These settings can be adjusted in the bot settings of Botpress (or your chosen tool), and there are many different models to choose from.

свернуть