import os import dotenv import streamlit as st from langchain import SQLDatabase from langchain.agents import OpenAIFunctionsAgent, AgentExecutor from langchain.agents.agent_toolkits import SQLDatabaseToolkit from langchain.agents.agent_toolkits.sql.prompt import SQL_PREFIX, SQL_FUNCTIONS_SUFFIX from langchain.agents.openai_functions_agent.agent_token_buffer_memory import ( AgentTokenBufferMemory, ) from langchain.callbacks import StreamlitCallbackHandler, LLMonitorCallbackHandler from langchain.chat_models import ChatOpenAI from langchain.memory.chat_message_histories import StreamlitChatMessageHistory from langchain.prompts import ( HumanMessagePromptTemplate, MessagesPlaceholder, ChatPromptTemplate, ) from langchain.schema import SystemMessage, AIMessage dotenv.load_dotenv() st.set_page_config( page_title="ChatSQL", page_icon="🦜", layout="wide", ) st.title("Chat with a SQL database") class MODELS: GPT_35 = "gpt-3.5-turbo" GPT_4 = "gpt-4" openai_api_key = os.getenv("OPENAI_API_KEY", None) with st.sidebar: openai_api_key = st.text_input("OpenAI API Key:", os.getenv("OPENAI_API_KEY", "")) model = st.selectbox("Model", (MODELS.GPT_35, MODELS.GPT_4)) if not openai_api_key: st.warning( "Please provide an OpenAI API key. You can get one [here](https://beta.openai.com/)." ) else: llm = ChatOpenAI(temperature=0, model_name=model, openai_api_key=openai_api_key) db = SQLDatabase.from_uri("sqlite:///Chinook.db") toolkit = SQLDatabaseToolkit(db=db, llm=llm) tools = toolkit.get_tools() prefix = SQL_PREFIX.format(dialect=toolkit.dialect, top_k=10) max_token_limit = 2000 messages = [ SystemMessage(content=prefix), MessagesPlaceholder(variable_name="history"), HumanMessagePromptTemplate.from_template("{input}"), AIMessage(content=SQL_FUNCTIONS_SUFFIX), MessagesPlaceholder(variable_name="agent_scratchpad"), ] prompt = ChatPromptTemplate(messages=messages) agent = OpenAIFunctionsAgent(llm=llm, tools=tools, prompt=prompt) msgs = StreamlitChatMessageHistory() memory = AgentTokenBufferMemory(llm=llm, chat_memory=msgs) agent_executor = AgentExecutor( agent=agent, tools=tools, memory=memory, verbose=True, return_intermediate_steps=True, ) monitor_callback = LLMonitorCallbackHandler() starter_message = "Ask me anything about the [Chinook digital media store](https://github.com/lerocha/chinook-database)!" if not msgs.messages or st.sidebar.button("Clear message history"): st.chat_message("assistant").write(starter_message) for msg in msgs.messages: if msg.type in ["human", "ai"] and msg.content: st.chat_message(msg.type).write(msg.content) if prompt := st.chat_input("How many artists are there?"): st.chat_message("user").write(prompt) st_callback = StreamlitCallbackHandler(st.container()) response = agent_executor( {"input": prompt, "history": msgs.messages}, callbacks=[st_callback, monitor_callback], include_run_info=True, ) st.chat_message("assistant").write(response["output"])