utils_codey.py

# Copyright 2023 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


"""
Utility module for Codey releated demo.
"""

from functools import partial
import pandas as pd
import streamlit as st
import tomllib
from typing import Optional

from google.cloud import bigquery
from google.cloud import datacatalog_v1
from google.cloud import translate_v2 as translate
from pandas import DataFrame
from utils_streamlit import reset_page_state
from vertexai.preview.language_models import TextGenerationModel


# Load configuration file
with open("./app_config.toml", "rb") as f:
    data = tomllib.load(f)

TEXT_MODEL_NAME = data["models"]["text"]["text_model_name"]
TRANSLATE_LANGUAGES = data["translate_api"]

translate_client = translate.Client()
prompt = data["pages"]["3_audiences"]["prompt_nl_sql"]


def get_tags_from_table(
        dataset_id: str, 
        table_id: str, 
        project_id: str, 
        tag_template_name: str
):
    """Gets the tags from a BigQuery table.

    Args:
        dataset_id:
             The ID of the BigQuery dataset that contains the table.
        table_id: 
            The ID of the BigQuery table.
        project_id: 
            The ID of the Google Cloud project.
        tag_template_name: 
            The name of the tag template.

    Returns:
        A string containing the tags for the table.
    """
    # Lookup Data Catalog's Entry referring to the table.
    datacatalog_client = datacatalog_v1.DataCatalogClient()
    resource_name = (
        f"//bigquery.googleapis.com/projects/{project_id}"
        f"/datasets/{dataset_id}/tables/{table_id}"
    )
    table_entry = datacatalog_client.lookup_entry(
        request={"linked_resource": resource_name}
    )

    # Make the request
    page_result = datacatalog_client.list_tags(parent=table_entry.name)
    tags_str = ''
    # Handle the response
    for response in page_result:
        if response.template == tag_template_name:
            desc = response.fields["description"].string_value
            data_type = response.fields["data_type"].string_value
            pk = response.fields["is_primary_key"].bool_value
            fk = response.fields["is_foreign_key"].bool_value            
            tags_str += ("Full table name: {} "
                         "- Column: {} " 
                         "- Data Type: {} " 
                         "- Primary Key: {} " 
                         "- Foreign Key: {} " 
                         "- Description: {}\n".format(
                f'`{project_id}.{dataset_id}.{table_id}`', 
                response.column, 
                data_type, 
                pk, 
                fk, 
                desc))

    return tags_str


def get_metadata_from_dataset(
        query: str,
        project_id: str,
        dataset_id: str, 
        tag_template_name: str,
        state_key: str
):
    """Gets the metadata for all tables in a BigQuery dataset.

    Args:
        query: 
            The BigQuery query to run to get the list of tables.
        project_id: 
            The ID of the BigQuery project.
        dataset_id: 
            The ID of the BigQuery dataset.
        tag_template_name: 
            The name of the BigQuery tag template to use to get the table descriptions.
        state_key: 
            The key to use to store the metadata in the Streamlit session state.
    """
    
    if len(st.session_state.get(state_key, [])) == 0:
        # print("Gets the metadata once")
        bqclient = bigquery.Client(project=project_id)
        query_job = bqclient.query(query)  # API request
        rows = query_job.result()
        metadata = []
    
        for row in rows:
            full_table_path = f'`{project_id}.{dataset_id}.{row.table_name}`'
            table_metadata = f'[SCHEMA details for table {full_table_path}]\n'

            table_metadata += get_tags_from_table(
                dataset_id, row.table_name, project_id, tag_template_name)
            metadata.append(table_metadata)

        st.session_state[state_key] = metadata


def get_full_context_from_list(metadata: list):
    """Gets the full context from a list of metadata.

    Args:
        metadata: A list of metadata objects.

    Returns:
        A string containing the full context.
    """
    context = ''
    for i in metadata:
        context += i
    return context


def generate_prompt(
        question: str,
        metadata: list,
        state_key: str
):
    """Generates a prompt for a GoogleSQL query compatible with BigQuery.

    Args:
        question: 
            The question to answer.
        metadata: 
            A list of dictionaries, where each dictionary describes a BigQuery table. 
            The dictionaries should have the following keys:
            - name: The name of the table.
            - schema: The schema of the table.
            - description: A description of the table.
        state_key: 
            The key to use to store the prompt in the session state.

    Returns:
        The prompt.
    """
    context = ''
    for i in metadata:
        context += i

    st.session_state[state_key] = f"""{prompt}
{context}
[Q]: {question}
[SQL]: 
"""


def generate_sql_and_query(
        state_key: str,
        title: str,
        query: str,
        project_id: str,
        dataset_id: str,
        tag_template_name: str,
        bqclient: bigquery.Client,
        prompt_example: str,
        fallback_query: str="",
        email_prompt_example: str="",
        text_model: TextGenerationModel=None) -> Optional[DataFrame]:
    """Generates a GoogleSQL query and executes it against a BigQuery dataset.

    Args:
        state_key: 
            A unique identifier for the current session.
        title: 
            The title of the UI page.
        query: 
            The initial query text.
        project_id: 
            The ID of the BigQuery project.
        dataset_id: 
            The ID of the BigQuery dataset.
        tag_template_name: 
            The name of the tag template to use for the query.
        bqclient: 
            A BigQuery client object.
        prompt_example: 
            An example prompt for the query.
        email_prompt_example: 
            An example prompt for the email.
        text_model: 
            A TextGenerationModel object.

    Returns:
        A DataFrame containing the results of the query.

    Raises:
        NotFoundError: If the dataset or table is not found.
        BadRequestError: If the query is invalid.
    """
    with st.form(f"{state_key}_form"):
        st.write(f"**{title}**")

        question = st.text_area(
            'Ask a question using natural language to the dataset above',
            key=f"{state_key}_question_prompt_text_area",
            value=prompt_example, height=50)

        # Every form must have a submit button.
        submit_button = st.form_submit_button("Submit")

    if submit_button:
        reset_page_state(state_key)
        if not question:
            st.warning("Question is empty")
            return None
        with st.spinner('Retrieving metadata from BigQuery'):
            get_metadata_from_dataset(
                query=query, 
                project_id=project_id, 
                dataset_id=dataset_id,
                tag_template_name=tag_template_name,
                state_key=f"{state_key}_Dataset_Metadata")
        
        with st.spinner('Creating a prompt'):
            generate_prompt(
                question, 
                st.session_state[f"{state_key}_Dataset_Metadata"],
                f"{state_key}_Prompt_Template")

        with st.expander('Prompt'):
            st.text(st.session_state[f"{state_key}_Prompt_Template"])
        
        with st.spinner('Generating the GoogleSQL statement with PaLM'):
            client_code_model = TextGenerationModel.from_pretrained(TEXT_MODEL_NAME)
            try:
                gen_code = client_code_model.predict(
                    prompt = st.session_state[f"{state_key}_Prompt_Template"],
                    max_output_tokens = 1024,
                    temperature=0.2
                )
            except:
                st.session_state[f"{state_key}_Gen_Code"] = fallback_query
            else:
                if gen_code and gen_code.text:
                    st.session_state[f"{state_key}_Gen_Code"] = gen_code.text
                else:
                    st.session_state[f"{state_key}_Gen_Code"] = fallback_query

        try:
            with st.spinner('Querying BigQuery...'):
                result_query = bqclient.query(st.session_state[f"{state_key}_Gen_Code"])
                result_query.result()
        except:
            st.session_state[f"{state_key}_Gen_Code"] = fallback_query
            with st.spinner('Querying BigQuery...'):
                try:
                    result_query = bqclient.query(st.session_state[f"{state_key}_Gen_Code"])
                    result_query.result()
                except:
                    if state_key == "TalkToData_insight":
                        st.session_state[f"{state_key}_Result_Final_Query"] = pd.DataFrame(
                            data["pages"]["3_audiences"]["example_result_0"])
                    elif state_key == "TalkToData_audience":
                        st.session_state[f"{state_key}_Result_Final_Query"] = pd.DataFrame(
                            data["pages"]["3_audiences"]["example_result_1"])
                    elif state_key == "TalkToData_followup":
                        st.session_state[f"{state_key}_Result_Final_Query"] = pd.DataFrame(
                            data["pages"]["3_audiences"]["example_result_2"])
                else:
                    st.session_state[
                        f"{state_key}_Result_Final_Query"] = result_query.to_dataframe()
    
            st.write('Resulting query generated by PaLM 2')
            st.write(f"""```sql
                    {st.session_state[f"{state_key}_Gen_Code"]}""")
            st.success('Query executed successfully. Retrieving dataset.')
            st.write('')
            st.write('Resulting table (limited by 50 rows)')
            st.dataframe(
                st.session_state[f"{state_key}_Result_Final_Query"].iloc[:50])
        else:
            st.write('Resulting query generated by PaLM 2')
            st.write(f"""```sql
                    {st.session_state[f"{state_key}_Gen_Code"]}""")
            st.success('Query executed successfully. Retrieving dataset.')
            st.session_state[
                f"{state_key}_Result_Final_Query"] = result_query.to_dataframe()
            st.write('')
            st.write('Resulting table (limited by 50 rows)')
            st.dataframe(
                st.session_state[f"{state_key}_Result_Final_Query"].iloc[:50])

    else:
        if f"{state_key}_Prompt_Template" in st.session_state:
            with st.expander('Prompt'):
                st.text(st.session_state[f"{state_key}_Prompt_Template"])

        if f"{state_key}_Gen_Code" in st.session_state:
            st.write('Resulting query generated by PaLM 2')
            st.write(st.session_state[f"{state_key}_Gen_Code"])

        if f"{state_key}_Result_Final_Query" in st.session_state:
            st.write('Resulting table (limited by 50 rows)')
            st.dataframe(
                st.session_state[f"{state_key}_Result_Final_Query"].iloc[:50])
    
    if f"{state_key}_Result_Final_Query" in st.session_state:
        if email_prompt_example:
            results_length = len(
                st.session_state[f"{state_key}_Result_Final_Query"])
            with st.form(f"{state_key}_email_form"):
                st.write(f"**Write an email**")

                col1, col2 = st.columns([70,30])

                with col1:
                    email_prompt = st.text_area(
                        'Write an email using natural language to each customer',
                        key=f"{state_key}_email_prompt_text_area",
                        value=email_prompt_example,
                        height=170)
                    number_emails = st.slider(
                        "Number of emails to generate",
                        min_value=1,
                        max_value=min(results_length, 500),
                        step=1,
                        value=int(min(results_length/2, 2)))
                
                with col2:
                    st.write('**Model parameters**')
                    temperature = st.slider('Temperature', 0.0, 1.0, 0.2)

                # Every form must have a submit button.
                email_submit_button = st.form_submit_button("Submit")

            if email_submit_button:
                if not email_prompt:
                    st.warning("Email prompt is empty.")
                    return None
                
                emails = {"email":[], "generated_email": []}

                bar = st.progress(0.0, f"Generating emails...0.0%")
                
                for i, row in st.session_state[
                        f"{state_key}_Result_Final_Query"].iloc[
                            :number_emails].iterrows():
                    

                    data_string = "\n".join(
                        [f"{row_key}: {row[row_key]}" for row_key in row.keys()])
                    
                    response = text_model.predict(
                            prompt=f"{data_string} \n {email_prompt}",
                            temperature=temperature,
                            max_output_tokens=1024,
                        ).text
                    
                    emails["email"].append(row.get("email","")),
                    emails["generated_email"].append(response)
                    
                    percent = (i+1)/number_emails
                    bar.progress(percent, f"Generating emails...{percent*100:.1f}%")
                
                st.success("Emails generated")
                
                st.session_state[f"{state_key}_Generated_Emails"] = DataFrame(
                    emails)
                st.session_state[
                    f"{state_key}_Generated_Emails_CSV"] = DataFrame(
                    emails).to_csv().encode('utf-8')
            
            if f"{state_key}_Generated_Emails" in st.session_state:
                st.write('Generated emails')
                st.session_state[f"{state_key}_Generated_Emails"] = st.data_editor(
                    st.session_state[f"{state_key}_Generated_Emails"]
                )
                st.download_button(
                    "Download csv", 
                    data=st.session_state[f"{state_key}_Generated_Emails_CSV"],
                    file_name="emails.csv",
                    mime="text/csv") 
                
                with st.form(key=f'{state_key}_translate_form'):
                    st.write(f"**Translate generated text**")

                    target_language_name = st.selectbox(
                        "Languages", options=TRANSLATE_LANGUAGES.keys())

                    translate_submit_button = st.form_submit_button(
                        label='Translate')

                if translate_submit_button:
                    translated_df = st.session_state[
                        f"{state_key}_Generated_Emails"].copy()
                    
                    with st.spinner("Translating..."):
                        translate_map = partial(
                            translate_client.translate,
                            source_language="en",
                            target_language=TRANSLATE_LANGUAGES[target_language_name])
                        translated_df["generated_email"] = translated_df[
                            "generated_email"].apply(
                            translate_map).apply(
                            lambda x: x.get("translatedText",""))
                        st.session_state[
                            f"{state_key}_Translated"] = translated_df
                        st.session_state[
                            f"{state_key}_Translated_CSV"
                            ] = translated_df.to_csv().encode('utf-8')
                    st.success("Emails translated")
                
                if f"{state_key}_Translated" in st.session_state:
                    st.write('Translated Generated emails')
                    st.session_state[f"{state_key}_Translated"] = st.data_editor(
                        st.session_state[f"{state_key}_Translated"]) 
                    st.download_button(
                        "Download CSV", 
                        data=st.session_state[f"{state_key}_Translated_CSV"],
                        file_name="translated_emails.csv",
                        mime="text/csv") 
  
        else:
            return st.session_state[f"{state_key}_Result_Final_Query"]
    
    return None