How to stream LLM response with streamlit?

[fabmeyer]

I am following this script using RetrievalQA chain.

Code:

llm = OpenAI(client=OpenAI, streaming=True, callbacks=[StreamingStdOutCallbackHandler()], temperature=0)
chain = RetrievalQA.from_chain_type(llm=llm, chain_type='refine', retriever=docsearch.as_retriever())

...

if 'user_input' not in st.session_state:
    st.session_state['user_input'] = []
if 'generated_text' not in st.session_state:
    st.session_state['generated_text'] = []

user_input = st.text_area('Enter a question', value=f"What are trends in {st.session_state['thematic']['term']}?")

button_2 = st.button('Get answer')

if user_input and button_2:
    st.session_state.user_input.append(user_input)
    with st.spinner('Running LLM...'):
        st.session_state.generated_text.append(st.session_state['chain'].run(user_input))

if 'generated_text' in st.session_state and len(st.session_state['generated_text']) > 0:
    for i in range(len(st.session_state['generated_text']) - 1, -1, -1):
        message(st.session_state['user_input'][i], is_user=True, key=str(i) + '_user')
        message(st.session_state['generated_text'][i], key=str(i))

How can I stream the response of the LLM in real time (like on the console)?

[tractorjuice]

Take a look at the new Streamlit Chat elements. It may help?

https://docs.streamlit.io/knowledge-base/tutorials/build-conversational-apps

[ucola]

@fabmeyer did you solve your problem?

[tractorjuice]

Use these instead: https://docs.streamlit.io/library/api-reference/chat

[tractorjuice]

Examples here, https://docs.streamlit.io/knowledge-base/tutorials/build-conversational-apps#build-a-chatgpt-like-app

[ucola]

@tractorjuice Yes, i read this implementation, but do not bring them up running. Streamlit write:
StreamlitAPIException: st.stream_write expects a generator or stream-like object as input not <class 'str'>. Please use st.write instead for this data type.

My code is bellow, did you have any idea how to rewrite this?

def prepare_llm(prompt):
    st.chat_message(name="user", avatar=IMAGE["user"]).markdown(prompt)
    st.session_state.messages.append({"role": "user", "content": prompt})

    msg = []
    msg.append({"role": "assistant", "content": description})
    for x in st.session_state.messages:
        msg.append(x)

    embeddings = OpenAIEmbeddings()
    docsearch = Pinecone.from_existing_index(
        index_name="langchain-index", embedding=embeddings
    )

    with (st.chat_message(name="assistant", avatar=IMAGE["assistant"])):
        llm = ChatOpenAI(
            streaming=True,
            callbacks=[StreamingStdOutCallbackHandler()],
            temperature=0,
        )
        qa = RetrievalQA.from_chain_type(
            llm=llm,
            chain_type='stuff',
            retriever=docsearch.as_retriever(),
            chain_type_kwargs={
                "prompt": prompt,
                "memory": ConversationBufferMemory(
                    memory_key="history",
                    input_key="question"),
            },
        )
        response = st.write_stream(
            qa.run({"query": "What Atlas Client?"})
        )

[Bharathi-A7]

@ucola I'm facing the same error. Were you able to resolve it?

[truevis]

This is working for me (with groq):

    full_response = ""  # Initialize outside the generator

    def generate_responses(completion):
        global full_response
        for chunk in completion:
            response = chunk.choices[0].delta.content or ""
            if response:
                full_response += response  # Append to the full response
                yield response
    st.chat_message("assistant")
    stream = generate_responses(completion)
    st.write_stream(stream)

    # After streaming
    if full_response:  # Check and use the full_response as needed
        response_message = {"role": "assistant", "content": full_response}
        # with st.chat_message("assistant"):
        #     st.markdown(full_response)
        st.session_state.messages.append(response_message)