ChatBGC

[matinnuhamunada]

With the advancement of better LLMs, it is now possible to interact with BGCFlow results through chatbot-like interface. This is possible by converting natural language into SQL queries.

Option to interact with BGCFlow duckdb result with LLMs:

1. Directly use Metabot feature from Metabase. This utilizes the OpenAI API key, and can be further used to create Slack Apps. Interaction can happen in the Metabase directly or through Slackbot. The downside is that the open source version feature is quite limited. Reference: https://discourse.metabase.com/t/metabot-openai/24660/31. Steps:

• Get an API key with Paid plan
• Configure the Admin setting for metabot (I used gpt-3.5-turbo)
• Create a Question with the raw data source
• Make sure you look once to the table for the column names
• save it
• Make it into a model
• Analyze the Model, i.e (it's columns, data points ... just take a glance)
• Ask question to the bot e.g: if a column stores staked price for each record and you want to find out the total staked amount. You can ask the bot that "Accumulate me the total staked price from the table" or so ...
• This gives me result other than the bot server busy
• Bonus TIP: You can ask complex questions to the bot once you know your data well

Direct interaction with DuckDB

Direct interaction with the DuckDB database can be achieved using Langchain and other stacks. This gives flexibility to run open source LLMs such as the latest Mistral model. Example project: https://github.com/kyaukyuai/chatDuck. There are many experimental open repository out there, but it uses this stacks:

• Langchain
• Ollama
• ChromaDB
• Streamlit

Challenges

While of course interacting with the database with natural language is convenient, this technology is still in its infancy. For SQL experts, it will be much faster to directly write queries than using the chatbots. Another big challenge is to include the context of the database schema into the model. I think this is doable by parsing the schema constraints and descriptions of the tables to the LLMs. Current schema and description are not optimal for this purpose yet: https://github.com/NBChub/bgcflow_dbt-duckdb/blob/main/models/raw/schema.yml

Another promising alternative

I think using knowledge graphs instead of SQL database is the way forward as it fits well with how the LLM works. An example project is https://github.com/neo4j/NaLLM