Best practice for sending CSV data to LLM for daily report analysis?

[Pichayut_Prasertwit]

Hello !

I’m working on a use case where I want to send daily sales reports (CSV format) to an LLM agent to generate summary and insights automatically.

Right now I’m considering two options:

1. Convert the CSV (from Google Sheets) into a single JSON object and send it directly to the LLM.
2. Upload the CSV into a vector store and let the LLM retrieve and analyze it from there.

My concern is that vector-based chunking might miss overall patterns or insights, while sending a large JSON might hit token limits. (I tried letting the LLM calculate the total sales, but it got the numbers wrong.

My goal is to get accurate, high-level analysis from these daily reports with minimal manual prep.

Would love to hear how others have handled similar cases. Any suggestions or alternatives are appreciated.

[mrrobot]

Hey @Pichayut_Prasertwit,

Why don’t you do the math yourself, send only the results to the LLM.
Calculate totals/averages in your preprocessing script
Send LLM a summary object with the numbers + a few sample rows
LLM writes the narrative, doesn’t calculate anything
This fixes the accuracy problem and avoids token limit

Ex

{
  "date": "2025-01-15",
  "total_sales": 45230.50,
  "vs_yesterday": "+12.3%",
  "top_products": [...],
  "key_metrics": {...},
  "sample_data": [10-20 most relevant rows]
} 

[cutecatcode]

I prefer the first options because I do like to calculate it first before I’m talking to AI about my document. Basically, I would like to count average, summary, and total about the sales, and also adding top 5 products if it’s possible.
If you’re using AI Agent in the n8n workflow, I would prefer to directly call the google sheet without converting it and talk with AI to do thing with my sheet