Chatbot for CSV data with calculations

Question

Chatbot for CSV data with calculations

Muhammad Zubair 30

I am looking for a chatbot that can respond to questions about CSV data, including making calculations on the data. I have added all of my CSV files as indexers in Cognitive Search, and the chatbot is responding correctly to questions about individual records. However, when I ask a general question, such as "What is the sum of all amounts from 2/2/2023 to 3/3/2023?", the chatbot does not respond correctly. It seems that the chatbot is only responding to questions about individual row, not to questions that require calculations on multiple rows or columns. Can anyone help me to achieve this?

Pramod Valavala 20,656 Reputation points Microsoft Employee

2023-10-19T15:12:33.22+00:00

@Muhammad Zubair Just following up here to see if the response here helps

1 answer

Your answer

Pramod Valavala 20,656 Reputation points Microsoft Employee

2023-10-19T15:12:33.22+00:00

@Muhammad Zubair Just following up here to see if the response here helps

Answer 1

Pramod Valavala 20,656 Microsoft Employee

@Muhammad Zubair Applying on LLMs on Tabular Data is a bit trickier compared to asking it questions about regular textual facts. In many cases, it would be impossible to pass in the complete tabular data into the prompt due to token limits, but it should work for data that can fit in the prompt.

In these cases, you would need to bring in all your data instead of using Azure Cognitive Search to fetch semantically relevant information, which is what Azure OpenAI on your Data does. Unfortunately, there is no way to change this behavior since this is by design.

Another approach would be to ingest aggregate information (like a pivot table in excel), which could be fetched during retrieval from Cognitive Search with the right prompt.

With more complex use cases like this where you might need to perform auxiliary autonomous tasks (like querying a database for example), it might be worthwhile to explore frameworks like Semantic Kernel which could help orchestrate and build more complex AI-powered workflows.

Muhammad Zubair 30 Reputation points

2023-10-12T20:14:34.9866667+00:00

Pramod Valavala, Thank you for your suggestions. What if we use function calling with Azure OpenAI Service for each field in every column? https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/function-calling

This way, we can pass the values and process them more effectively.

However, I've observed that when we add CSV data using the indexer in Azure Cognitive Search, it seems to perform searches in a row-wise manner rather than searching within the columns of the data.

I'd appreciate your insights on how we can address this issue and find a solution that aligns with our goals. Your guidance would be highly valued.

Thank you in advance for your assistance.
Pramod Valavala 20,656 Reputation points Microsoft Employee

2023-10-16T17:17:17.6366667+00:00
@Muhammad Zubair Yes. That is an option, like the approach with Semantic Kernel, but to fetch all the columns values instead of individual row, you would need to index documents with column data as well. Again, semantic search may not really do well with just a bunch of numbers in the document without enough context.

Instead, using the functions feature, you could define a function that would extract intent from the query which you could use to translate into a database query on your backend to run against the CSV data.

In langchain for example, there is a SQL Agent that translates input prompts into SQL queries to run against a database. Essentially, the idea is to use OpenAI to "write" the SQL query appropriate for your database.

For the sample input prompt that you've mentioned in your original post

What is the sum of all amounts from 2/2/2023 to 3/3/2023?

The expected output (intermediate output in the case of langchain or similar) would be something like this

SELECT SUM(amount) FROM transactions WHERE transaction_date >= '2023-02-02' AND transaction_date <= '2023-03-03';

Share via

Chatbot for CSV data with calculations

1 answer

Your answer