Pasting a big CSV into a chat and asking “what's the average?” is a recipe for wrong numbers. Models aren't calculators. The reliable pattern is:
- Code computes the numbers.
- The model interprets them and writes the explanation.
1. Profile the data with pandas
import pandas as pd
df = pd.read_csv("sales.csv", parse_dates=["date"])
profile = {
"rows": len(df),
"columns": {c: str(t) for c, t in df.dtypes.items()},
"by_month": df.groupby(df["date"].dt.to_period("M"))["revenue"].sum().round(2).astype(str).to_dict(),
"top_products": df.groupby("product")["revenue"].sum().nlargest(10).round(2).to_dict(),
"by_region": df.groupby("region")["revenue"].agg(["sum", "mean", "count"]).round(2).to_dict(),
}2. Ask for insights on the summary
import json
r = client.chat.completions.create(model="claude-sonnet-4-6", messages=[{"role": "user", "content":
"You are a data analyst. Using ONLY these computed figures, write: 3 key findings, 2 anomalies worth checking, "
"and 3 questions to investigate next. Quote numbers exactly as given.\n\n" + json.dumps(profile, default=str)}])
print(r.choices[0].message.content)3. Let the model write the analysis code
For follow-up questions, ask the model for pandas code instead of an answer, review it, and run it yourself:
Given a DataFrame `df` with columns {columns}, write pandas code that answers: "Which region grew fastest quarter over quarter?" Output only code.That way every number comes from your data, not from the model's guess.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.