Attention!

Since version 1.24.0 streamlit provides official elements to build conversational apps.

The new elements are more flexible, extensible and better supported, I would suggest to use them.

However, streamlit>=1.23 requires protobuf>=4 when some package requires protobuf<=3. In this condition you can use this package(<1.0.0) with streamlit<=1.22 as alternative. They are all simple to render text messages.

This package(>=1.0.0) will focus on wrapper of official chat elements to make chat with LLMs more convenient.

Chatbox component for streamlit

A Streamlit component to show chat messages. It's basiclly a wrapper of streamlit officeial elements including the chat elemnts.

• demo

• demo agent

Features

• support streaming output.
• support markdown/image/video/audio messages, and all streamlit elements could be supported by customized OutputElement.
• output multiple messages at once, and make them collapsable.
• maintain session state context bound to chat conversation
• export & import chat histories

This make it easy to chat with langchain LLMs in streamlit.

Goto webui of langchain-chatchat to see the actual application.

Install

just pip install -U streamlit-chatbox

Usage examples

import streamlit as st
from streamlit_chatbox import *
import time
import simplejson as json


llm = FakeLLM()
chat_box = ChatBox(
    use_rich_markdown=True, # use streamlit-markdown
    user_theme="green", # see streamlit_markdown.st_markdown for all available themes
    assistant_theme="blue",
)
chat_box.use_chat_name("chat1") # add a chat conversatoin

def on_chat_change():
    chat_box.use_chat_name(st.session_state["chat_name"])
    chat_box.context_to_session() # restore widget values to st.session_state when chat name changed


with st.sidebar:
    st.subheader('start to chat using streamlit')
    chat_name = st.selectbox("Chat Session:", ["default", "chat1"], key="chat_name", on_change=on_chat_change)
    chat_box.use_chat_name(chat_name)
    streaming = st.checkbox('streaming', key="streaming")
    in_expander = st.checkbox('show messages in expander', key="in_expander")
    show_history = st.checkbox('show session state', key="show_history")
    chat_box.context_from_session(exclude=["chat_name"]) # save widget values to chat context

    st.divider()

    btns = st.container()

    file = st.file_uploader(
        "chat history json",
        type=["json"]
    )

    if st.button("Load Json") and file:
        data = json.load(file)
        chat_box.from_dict(data)


chat_box.init_session()
chat_box.output_messages()

def on_feedback(
    feedback,
    chat_history_id: str = "",
    history_index: int = -1,
):
    reason = feedback["text"]
    score_int = chat_box.set_feedback(feedback=feedback, history_index=history_index) # convert emoji to integer
    # do something
    st.session_state["need_rerun"] = True


feedback_kwargs = {
    "feedback_type": "thumbs",
    "optional_text_label": "wellcome to feedback",
}

if query := st.chat_input('input your question here'):
    chat_box.user_say(query)
    if streaming:
        generator = llm.chat_stream(query)
        elements = chat_box.ai_say(
            [
                # you can use string for Markdown output if no other parameters provided
                Markdown("thinking", in_expander=in_expander,
                         expanded=True, title="answer"),
                Markdown("", in_expander=in_expander, title="references"),
            ]
        )
        time.sleep(1)
        text = ""
        for x, docs in generator:
            text += x
            chat_box.update_msg(text, element_index=0, streaming=True)
        # update the element without focus
        chat_box.update_msg(text, element_index=0, streaming=False, state="complete")
        chat_box.update_msg("\n\n".join(docs), element_index=1, streaming=False, state="complete")
        chat_history_id = "some id"
        chat_box.show_feedback(**feedback_kwargs,
                                key=chat_history_id,
                                on_submit=on_feedback,
                                kwargs={"chat_history_id": chat_history_id, "history_index": len(chat_box.history) - 1})
    else:
        text, docs = llm.chat(query)
        chat_box.ai_say(
            [
                Markdown(text, in_expander=in_expander,
                         expanded=True, title="answer"),
                Markdown("\n\n".join(docs), in_expander=in_expander,
                         title="references"),
            ]
        )

cols = st.columns(2)
if cols[0].button('show me the multimedia'):
    chat_box.ai_say(Image(
        'https://tse4-mm.cn.bing.net/th/id/OIP-C.cy76ifbr2oQPMEs2H82D-QHaEv?w=284&h=181&c=7&r=0&o=5&dpr=1.5&pid=1.7'))
    time.sleep(0.5)
    chat_box.ai_say(
        Video('https://sample-videos.com/video123/mp4/720/big_buck_bunny_720p_1mb.mp4'))
    time.sleep(0.5)
    chat_box.ai_say(
        Audio('https://sample-videos.com/video123/mp4/720/big_buck_bunny_720p_1mb.mp4'))

if cols[1].button('run agent'):
    chat_box.user_say('run agent')
    agent = FakeAgent()
    text = ""

    # streaming:
    chat_box.ai_say() # generate a blank placeholder to render messages
    for d in agent.run_stream():
        if d["type"] == "complete":
            chat_box.update_msg(expanded=False, state="complete")
            chat_box.insert_msg(d["llm_output"])
            break

        if d["status"] == 1:
            chat_box.update_msg(expanded=False, state="complete")
            text = ""
            chat_box.insert_msg(Markdown(text, title=d["text"], in_expander=True, expanded=True))
        elif d["status"] == 2:
            text += d["llm_output"]
            chat_box.update_msg(text, streaming=True)
        else:
            chat_box.update_msg(text, streaming=False)

btns.download_button(
    "Export Markdown",
    "".join(chat_box.export2md()),
    file_name=f"chat_history.md",
    mime="text/markdown",
)

btns.download_button(
    "Export Json",
    chat_box.to_json(),
    file_name="chat_history.json",
    mime="text/json",
)

if btns.button("clear history"):
    chat_box.init_session(clear=True)
    st.experimental_rerun()


if show_history:
    st.write(st.session_state)

Todos

• wrapper of official chat elements

• input messages: (this depends on the official st.chat_input improvement by #7069)

  • TEXT
  • IMAGE
    • file upload
    • paste from clipboard(streamlit_bokeh_events)
  • VIDEO
    • file upload
  • AUDIO
    • file upload
    • audio-recorder-streamlit

• output message types:

  • Text/Markdown/Image/Audio/Video
  • any other output types supported by streamlit

• improve output performance

  • streaming output message
  • show message in expander
  • rich output message using streamlit-markdown
  • feedback by user

• export & import chat history

  • export to markdown
  • export to json
  • import json

• support output of langchain' Agent.

• conext bound to chat

changelog

v1.1.13

• add Json output element
• can choose to use streamlit-markdown instead of st.markdown. currently need streamlit==1.37.1 when streaming
• user can register custom output method with ChatBox.register_output_method. This is useful to use thirdparty components:

  from streamlit_chatbox import *
  from streamlit_markdown import st_hack_markdown
  
  ChatBox.register_output_method("st_markdown", st_hack_markdown)
  cb = ChatBox()
  cb.user_say(OutputElement("user defined output method", output_method="st_markdown", theme_color="blue", mermaid_theme_CSS=""))