{
    "action": "tool_call",
    "reasoning": "My earlier draft cited quality issues without any query results behind it. I'll profile the raw table in one query: nulls, distinct counts and duplicate User_IDs, plus ranges on the numeric columns.",
    "tool": "query_sql",
    "args": {
        "sql": "SELECT COUNT(*) AS n_rows, COUNT(DISTINCT User_ID) AS distinct_ids, COUNT(*) - COUNT(DISTINCT User_ID) AS dup_ids, COUNT(DISTINCT Gender) AS gender_variants, COUNT(DISTINCT Education_Level) AS edu_variants, COUNT(DISTINCT AI_Tool) AS tool_variants, COUNT(DISTINCT AI_Purpose) AS purpose_variants, COUNT(DISTINCT Would_Recommend) AS recommend_variants, COUNT(DISTINCT User_Type) AS usertype_variants, SUM(CASE WHEN Age IS NULL OR TRIM(Age) = '' THEN 1 ELSE 0 END) AS age_blank, SUM(CASE WHEN Gender IS NULL OR TRIM(Gender) = '' THEN 1 ELSE 0 END) AS gender_blank, SUM(CASE WHEN Monthly_Income IS NULL OR TRIM(Monthly_Income) = '' THEN 1 ELSE 0 END) AS income_blank, SUM(CASE WHEN Productivity_Score IS NULL THEN 1 ELSE 0 END) AS prod_null, MIN(Productivity_Score) AS prod_min, MAX(Productivity_Score) AS prod_max, MIN(Accuracy_Rating) AS acc_min, MAX(Accuracy_Rating) AS acc_max, MIN(Satisfaction_Score) AS sat_min, MAX(Satisfaction_Score) AS sat_max, MIN(Age) AS age_min_text, MAX(Age) AS age_max_text FROM raw.AI_Usage_and_Impact_on_Students_and_Professionals",
        "purpose": "Profile data quality: duplicates, blanks, category variant counts, numeric ranges",
        "source": "raw"
    }
}