Issues in the current project

[parthasarathydNU]

Areas of Improvement

1. We have the base foundation to do querying across tables

   • Halucinates - Even though the sql query output is empty, it returns a hallucinated response
   • Issues with classification - a given user message, sometimes, get's classified as query
   • Sometimes the generated query is wrong - fails during the query execution

2. Chart generation

   • Only one table's sample data is hardcoded in the UI side, even we are passing the same table name to the chain, so it can get the schema, so the charts are getting generated fine, but it is only restricted to the same table
   • For plotting the chart, we need both the data and the chart spec ( generated by llm ) - How do we get the right data for every user query while generating the chart
   • We don't know if the generated chart is the right one
   • More like an implementation problem

3. Usability - not very intuitive

4. We need the right users for feedback

5. Hallucinating while having a normal conversation

6. Final response generation after generating the SQL query and results is slow

7. Sometimes when asked questions out of scope, in a repetitive manner, the responses look very hard coded , it gives the same response over and over again

8. Too complex a backend - need to simplify it, too many chains

9. The chatbot seems to be responding to the previous query rather than the current query