Project
About project
About project
About project
The experience and knowledge shared by top specialists at IT conferences often remain only in the memory of attendees — or locked away in hours-long YouTube recordings. If a talk never gets turned into an article, the information is essentially lost and nearly impossible to retrieve later.
To preserve and make the best industry content accessible, Doubletapp’s ML engineers created a service trained on agency conference materials. It can answer almost any question in the fields of management, finance, business process organization, marketing, and sales — drawing on the expertise of recognized professionals.
How do you structure a marketing department in an IT company? What roles make up the ideal sales team? How can a CEO learn to delegate tasks? For all of this, you can ask Digital Guru.
How the Bot Works
How the Bot Works
How the Bot Works
- Register in the Telegram bot.
- Ask a question and wait a few seconds.
- You’ll see a text-based answer along with a link to the video it was sourced from.
- Click the link to watch the full talk.
Result
Result
Result
The project was successfully presented at the AGD IT conference in Yekaterinburg. Digital Guru helped participants quickly find information on topics from previous conference talks, revisit earlier presentations by the same speakers, and improve networking efficiency.
17.6% of attendees used the service, submitting a total of 121 questions. By presenting our ML tools at industry events, we aim to raise awareness around large language models and position ourselves as experts in this field.
Development
Python
ChatGPT
Chroma DB
LangChain
Our super team
FedorBackend
PavelDesignTechnical Implementation
Technical Implementation
Technical Implementation
The system is built using Retrieval-Augmented Generation (RAG), which is a technology that retrieves relevant information from a knowledge base and uses a language model to generate answers. [Learn more about RAG here]
How does Digital Guru find answers in its knowledge base?
- We collected several hundred video recordings of talks from the last two years.
- Transcribed the talks and split them into small chunks.
- Processed the chunks with a model that extracts semantic features and stores them in a vector database. This model is called an embedder, and the feature vector is an embedding.
- Packaged the service into a web interface and a Telegram bot.
- A user can now send a question to the bot in chat.
- The question is converted into a semantic vector (embedding).
- The system finds the chunks of information closest to the question’s vector.
- It selects five talks with the most relevant segments.
- A prompt is sent to the LLM, which, in simplified form, looks like this: “Here is the question, here is the context where the answer should be found. Search and provide the answer.”
- The LLM generates an answer and we send it to the user.
