How to display streaming response chunks from Azure OpenAI's LLM in Langchain on Gradio's new Chat-Interface

import os

import gradio as gr
import openai
from langchain.vectorstores import Chroma
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.chat_models import AzureChatOpenAI
from langchain.memory import ConversationBufferWindowMemory

from langchain.callbacks.manager import CallbackManager
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
from langchain.prompts import PromptTemplate
from langchain.chains import ConversationalRetrievalChain

from named_tuples import ConversationalRetrievalChainResponse

openai.api_base = os.getenv("OPENAI_API_BASE")
openai.api_key = os.getenv("OPENAI_API_KEY")
openai.api_type = "azure"
openai.api_version = os.getenv("OPENAI_API_VERSION")

# The LLM
llm = AzureChatOpenAI(

# The Prompt
template = """


QA_CHAIN_PROMPT = PromptTemplate.from_template(template)

# The Memory
buffer_window_memory = ConversationBufferWindowMemory(
    memory_key="chat_history", return_messages=True, k=4

# The Embeddings
openai_embeddings = OpenAIEmbeddings(

# The Knowledge Base
persist_directory = "vector_database/chromadb/"
vectordb = Chroma(
    persist_directory=persist_directory, embedding_function=openai_embeddings

# The Retriever
retriever = vectordb.as_retriever(search_type="mmr")

# The chain
chain = ConversationalRetrievalChain.from_llm(
    combine_docs_chain_kwargs={"prompt": QA_CHAIN_PROMPT},

# The function that takes a message as an argument and returns a streaming LLM resposne
def ask_conversational_retrieval_chain(
) -> ConversationalRetrievalChainResponse:
    result = chain({"question": question})
    return ConversationalRetrievalChainResponse(result=result["answer"])

# The Gradio App
def respond(message, history):
    return str(ask_conversational_retrieval_chain(message).result)

with gr.Blocks(title="Chatbot") as demo:
    gr.ChatInterface(respond, autofocus=True)

if __name__ == "__main__":
    demo.queue(concurrency_count=5, max_size=20).launch()

The function: ask_conversational_retrieval_chain generates a streaming response in the form of chunks and I would like to print those chunks as they’re being generated by the LLM directly to the chat interface.

I’ve came across a few methods for streaming but those methods wait for the whole response to be generated before giving a “streaming like” response.

P.S: I am new to all this, so any help would be appreciated!

Hi @sqy-harshil the basic idea would be to concatenate those chunks as they come in and simply yield the concatenated chunks instead of returning in this function:

def respond(message, history):
    return str(ask_conversational_retrieval_chain(message).result)

Here’s a simplified but working example:

import time
import gradio as gr

def slow_echo(message, history):
    for i in range(len(message)):
        yield "You typed: " + message[: i+1]

demo = gr.ChatInterface(slow_echo).queue()

if __name__ == "__main__":

Hi @sqy-harshil, just wondering if you managed to solve it? My workflow is similar - using ConversationalRetrievalChain, but I’m unable to find a way to stream the response.

I also tried streaming streaming the responses through the yield chunk method, but it actually first collects the entire response then starts sending in chunks through yield, which is actually worse. I tried many other things as well, now I am going to approach this with threads.
@abidlabs Do you have any working example with openai llm and langchain in python? It would help a lot. Thanks.

Hi @chitiii yes, please take a look here: Creating A Chatbot Fast

@DesmondChoy Hi,
I was able to resolve it using Langchain’s BaseCallbackHandler in streamlit,

from langchain.callbacks.base import BaseCallbackHandler

class StreamHandler(BaseCallbackHandler):
    def __init__(self, container, initial_text=""):
        self.container = container
        self.text = initial_text

    def on_llm_new_token(self, token: str, **kwargs):
        self.text += token

from streaming import StreamHandler

# The logic for generating the ConversationalRetrievalChain.from_llm...

callback = StreamHandler(st.empty())
response =, callbacks=callback)

# Rest of the code

This should give you a streaming response in chunks as the LLM is generating them.
You could either implement this in Gradio as per your use case or refer to
Gradio’s official documentation for how to achieve a streaming response from LLMs

1 Like