Census GPT

Aquest web contesta preguntes sobre el cens dels EUA utilitzant ChatGPT. Com ho fa? La clau està en la funció make_default_messages del fitxer text_to_sql.py, que explica a ChatGPT l’estructura de la base de dades del cens. A partir d’això i d’uns pocs exemples, ChatGPT pot traduir les preguntes subsegüents de l’usuari a sentències SQL que el web executarà i processarà.

def make_default_messages(table_names: List[str]):
  return [
    {
      "role": "system",
      "content": (
        "You are a helpful assistant for generating syntactically correct read-only SQL to answer a given question or command, generally about crime, demographics, and population."
        "\n"
        "The following are schemas of tables you can query:\n"
        "---------------------\n" + generate_msg_with_schemas(table_names) +
        "\n\n"
        "---------------------\n"
        "Use state abbreviations for states."
        " Table ’crime_by_city’ does not have columns ’zip_code’ or ’county’."
        " Do not use ambiguous column names."
        " For example, ’city’ can be ambiguous because both tables ’acs_census_data’ and ’crime_by_city’ have a column named ’city’."
        " Always specify the table where you are using the column."
        " If you include a ’city’ column in the result table, include a ’state’ column too."
        " If you include a ’county’ column in the result table, include a ’state’ column too."
        " Make sure each value in the result table is not null.\n"
      )
    },

etc.