We created a WikiBot service that learns from documentation and advises users in chat as a first-line support specialist. Our chatbot understands the question and answers like a human.
The whole world is moving towards the use of AI assistants. People like to get results just by giving commands in their native language.
I suggest you, together with my team, create a Dashboard.AI product that will have the following functionality:
- Search reports. The manager can ask the bot to find the required report. For example: “Find a report on advertising expenses for last year.” The bot, on the list in the data catalog, confluence, sends a link to the chat.
- Writing SQL queries. Instead of searching for the required report, the manager writes in the chat to the bot: “profit in Moscow for electronics from May to July.” The bot sends the necessary indicators and links to the desired report with filters installed. If there is no suitable report, then it writes an SQL query.
- Automatic creation of dashboards. The manager can ask the chatbot to create a dashboard. The bot, after a few clarifying questions, will select the optimal template and collect all the necessary SQL queries and filters in it.
Implementation
Search reports. The list of reports is usually located in a knowledge base, documentation, or data catalog. Modern products such as SiteGPT, ChatBase, WonderChat, and Wikibot (from now on referred to as the agent) are excellent at performing a semantic search:
- The documentation is indexed and stored in a vector database.
- When a user enters a query, the vector database contains the 5-10 closest vectors (semantic search). Each vector is the name and description of the report.
- The user’s request and the top related vectors are sent to LLM – a large language model (LLaMA, Falcon, Google Bard, Anthropic Claude), and it selects suitable reports for the user.
Writing SQL Queries
Modern LLMs already write SQL quite well ( an article comparing GPT, Claude, Bard ); to do this, they must ask a question and promptly pass a description of the tables. Having received a user request, the agent can take the names of the tables from the documentation and all the fields of the table from the DBMS metadata.
The following projects are currently gaining momentum:
- DB-GPT – allows you to generate database queries using local LLMs
- PandasAI – makes pandas (and other popular data analysis libraries) conversational by allowing you to ask questions of your Data Frames in natural language.
Automatic Creation Of Dashboards
GPT, having received the data table, can already select from it the most significant and suitable
columns for visualization. A comprehensive example of how to do this is described in the article “Create AI-powered dashboards.” By adding to this dashboard templates various best practices in the form of heuristics and clarifying queries from the bot, you can get a Junior BI developer.
A Look Into The Future
In the implementation section, I described what can be done now.
Here are a couple of ideas for the future:
- If you teach the “Dashboard.AI” project to create dashboards, then the next step will be to transfer dashboards from one system to another. Nowadays, a considerable amount of resources is required to transfer a company’s analytical reporting to an alternative BI product. In addition, it will be possible to create dashboards for several popular products at once (Tableau, Power BI, Superset, Metabase, Redash)
- A personal analyst who helps managers quickly obtain the necessary indicators and look for insights can become a super popular product.
Also Read: Limitations Of Current AI Models
