app.py

import os
import streamlit as st
from streamlit_chat import message as st_message
from sqlalchemy import create_engine

from langchain.agents import Tool, initialize_agent
from langchain.chains.conversation.memory import ConversationBufferMemory

from llama_index import GPTSQLStructStoreIndex, LLMPredictor, ServiceContext
from llama_index import SQLDatabase as llama_SQLDatabase
from llama_index.indices.struct_store import SQLContextContainerBuilder

from constants import (
    DEFAULT_SQL_PATH,
    DEFAULT_BUSINESS_TABLE_DESCRP,
    DEFAULT_VIOLATIONS_TABLE_DESCRP,
    DEFAULT_INSPECTIONS_TABLE_DESCRP,
    DEFAULT_LC_TOOL_DESCRP,
)
from utils import get_sql_index_tool, get_llm


@st.cache_resource
def initialize_index(
    llm_name, model_temperature, table_context_dict, api_key, sql_path=DEFAULT_SQL_PATH
):
    """Create the GPTSQLStructStoreIndex object."""
    llm = get_llm(llm_name, model_temperature, api_key)

    engine = create_engine(sql_path)
    sql_database = llama_SQLDatabase(engine)

    context_container = None
    if table_context_dict is not None:
        context_builder = SQLContextContainerBuilder(
            sql_database, context_dict=table_context_dict
        )
        context_container = context_builder.build_context_container()

    service_context = ServiceContext.from_defaults(llm_predictor=LLMPredictor(llm=llm))
    index = GPTSQLStructStoreIndex(
        [],
        sql_database=sql_database,
        sql_context_container=context_container,
        service_context=service_context,
    )

    return index


@st.cache_resource
def initialize_chain(llm_name, model_temperature, lc_descrp, api_key, _sql_index):
    """Create a (rather hacky) custom agent and sql_index tool."""
    sql_tool = Tool(
        name="SQL Index",
        func=get_sql_index_tool(
            _sql_index, _sql_index.sql_context_container.context_dict
        ),
        description=lc_descrp,
    )

    llm = get_llm(llm_name, model_temperature, api_key=api_key)

    memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)

    agent_chain = initialize_agent(
        [sql_tool],
        llm,
        agent="chat-conversational-react-description",
        verbose=True,
        memory=memory,
    )

    return agent_chain


st.title("🦙 Llama Index SQL Sandbox 🦙")
st.markdown(
    (
        "This sandbox uses a sqlite database by default, powered by [Llama Index](https://gpt-index.readthedocs.io/en/latest/index.html) ChatGPT, and LangChain.\n\n"
        "The database contains information on health violations and inspections at restaurants in San Francisco."
        "This data is spread across three tables - businesses, inspections, and violations.\n\n"
        "Using the setup page, you can adjust LLM settings, change the context for the SQL tables, and change the tool description for Langchain."
        "The other tabs will perform chatbot and text2sql operations.\n\n"
        "Read more about LlamaIndexes structured data support [here!](https://gpt-index.readthedocs.io/en/latest/guides/tutorials/sql_guide.html)"
    )
)


setup_tab, llama_tab, lc_tab = st.tabs(
    ["Setup", "Llama Index", "Langchain+Llama Index"]
)

with setup_tab:
    st.subheader("LLM Setup")
    api_key = st.text_input("Enter your OpenAI API key here", type="password")
    llm_name = st.selectbox(
        "Which LLM?", ["text-davinci-003", "gpt-3.5-turbo", "gpt-4"]
    )
    model_temperature = st.slider(
        "LLM Temperature", min_value=0.0, max_value=1.0, step=0.1
    )

    st.subheader("Table Setup")
    business_table_descrp = st.text_area(
        "Business table description", value=DEFAULT_BUSINESS_TABLE_DESCRP
    )
    violations_table_descrp = st.text_area(
        "Business table description", value=DEFAULT_VIOLATIONS_TABLE_DESCRP
    )
    inspections_table_descrp = st.text_area(
        "Business table description", value=DEFAULT_INSPECTIONS_TABLE_DESCRP
    )

    table_context_dict = {
        "businesses": business_table_descrp,
        "inspections": inspections_table_descrp,
        "violations": violations_table_descrp,
    }

    use_table_descrp = st.checkbox("Use table descriptions?", value=True)
    lc_descrp = st.text_area("LangChain Tool Description", value=DEFAULT_LC_TOOL_DESCRP)

with llama_tab:
    st.subheader("Text2SQL with Llama Index")
    if st.button("Initialize Index", key="init_index_1"):
        st.session_state["llama_index"] = initialize_index(
            llm_name,
            model_temperature,
            table_context_dict if use_table_descrp else None,
            api_key,
        )

    if "llama_index" in st.session_state:
        query_text = st.text_input(
            "Query:", value="Which restaurant has the most violations?"
        )
        use_nl = st.checkbox("Return natural language response?")
        if st.button("Run Query") and query_text:
            with st.spinner("Getting response..."):
                try:
                    response = st.session_state["llama_index"].as_query_engine(synthesize_response=use_nl).query(query_text)
                    response_text = str(response)
                    response_sql = response.extra_info["sql_query"]
                except Exception as e:
                    response_text = "Error running SQL Query."
                    response_sql = str(e)

            col1, col2 = st.columns(2)
            with col1:
                st.text("SQL Result:")
                st.markdown(response_text)

            with col2:
                st.text("SQL Query:")
                st.markdown(response_sql)

with lc_tab:
    st.subheader("Langchain + Llama Index SQL Demo")

    if st.button("Initialize Agent"):
        st.session_state["llama_index"] = initialize_index(
            llm_name,
            model_temperature,
            table_context_dict if use_table_descrp else None,
            api_key,
        )
        st.session_state["lc_agent"] = initialize_chain(
            llm_name,
            model_temperature,
            lc_descrp,
            api_key,
            st.session_state["llama_index"],
        )
        st.session_state["chat_history"] = []

    model_input = st.text_input(
        "Message:", value="Which restaurant has the most violations?"
    )
    if "lc_agent" in st.session_state and st.button("Send"):
        model_input = "User: " + model_input
        st.session_state["chat_history"].append(model_input)
        with st.spinner("Getting response..."):
            response = st.session_state["lc_agent"].run(input=model_input)
        st.session_state["chat_history"].append(response)

    if "chat_history" in st.session_state:
        for msg in st.session_state["chat_history"]:
            st_message(msg.split("User: ")[-1], is_user="User: " in msg)