import requests
import pandas as pd
OLLAMA_URL = "http://localhost:11434/api/generate"
MODEL = "llama3.2:1b"
def call_ollama(prompt: str) -> str:
"""Send a prompt to the local Ollama server."""
response = requests.post(
OLLAMA_URL,
json={
"model": MODEL,
"prompt": prompt,
"temperature": 0.1,
"stream": False,
},
)
response.raise_for_status()
return response.json()["response"].strip()
def summarize_outliers(df: pd.DataFrame) -> str:
"""Generate a technical summary and recommendation for detected outliers."""
if "outlier" not in df.columns:
return "Outlier analysis not found. Please run analyze_outliers first."
inliers = int((df["outlier"] == "inlier").sum())
outliers = int((df["outlier"] == "outlier").sum())
prompt = (
"You are a senior data analyst. The dataset was analyzed using "
"Isolation Forest.\n\n"
"Summary statistics:\n"
f"- Inliers detected: {inliers}\n"
f"- Outliers detected: {outliers}\n\n"
"Provide a technical interpretation of what this implies about the "
"dataset structure, potential data quality issues, or anomalies.\n\n"
"Then recommend a course of action regarding these outliers. Should they "
"be removed, investigated further, or used for model tuning? Justify your "
"recommendation."
)
return call_ollama(prompt)
def describe_dataset(df: pd.DataFrame) -> str:
"""Generate a technical summary of numeric and categorical distributions."""
numeric_desc = (
df.select_dtypes(include="number")
.describe()
.round(2)
.to_string()
)
categorical_cols = df.select_dtypes(include="object").columns
category_summary = (
"\n".join(
f"{col}: {dict(df[col].value_counts().head(5))}"
for col in categorical_cols
)
if not categorical_cols.empty
else "No categorical columns."
)
prompt = (
"You are a statistical analyst reviewing a structured dataset.\n\n"
f"Numerical Feature Summary:\n{numeric_desc}\n\n"
f"Categorical Feature Summary:\n{category_summary}\n\n"
"Please identify key patterns, distributions, and any potential data "
"skew or imbalance. Summarize these findings in a technical but concise "
"report suitable for inclusion in a data quality assessment."
)
return call_ollama(prompt)
