FastAPI StreamingResponse not streaming with generator function

Question:

I have a relatively simple FastAPI app that accepts a query and streams back the response from ChatGPT’s API. ChatGPT is streaming back the result and I can see this being printed to console as it comes in.

What’s not working is the StreamingResponse back via FastAPI. The response gets sent all together instead. I’m really at a loss as to why this isn’t working.

Here is the FastAPI app code:

import os
import time

import openai

import fastapi
from fastapi import Depends, HTTPException, status, Request
from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
from fastapi.responses import StreamingResponse

auth_scheme = HTTPBearer()
app = fastapi.FastAPI()

openai.api_key = os.environ["OPENAI_API_KEY"]


def ask_statesman(query: str):
    #prompt = router(query)
    
    completion_reason = None
    response = ""
    while not completion_reason or completion_reason == "length":
        openai_stream = openai.ChatCompletion.create(
            model="gpt-3.5-turbo",
            messages=[{"role": "user", "content": query}],
            temperature=0.0,
            stream=True,
        )
        for line in openai_stream:
            completion_reason = line["choices"][0]["finish_reason"]
            if "content" in line["choices"][0].delta:
                current_response = line["choices"][0].delta.content
                print(current_response)
                yield current_response
                time.sleep(0.25)


@app.post("/")
async def request_handler(auth_key: str, query: str):
    if auth_key != "123":
        raise HTTPException(
            status_code=status.HTTP_401_UNAUTHORIZED,
            detail="Invalid authentication credentials",
            headers={"WWW-Authenticate": auth_scheme.scheme_name},
        )
    else:
        stream_response = ask_statesman(query)
        return StreamingResponse(stream_response, media_type="text/plain")


if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000, debug=True, log_level="debug")

And here is the very simple test.py file to test this:

import requests

query = "How tall is the Eiffel tower?"
url = "http://localhost:8000"
params = {"auth_key": "123", "query": query}

response = requests.post(url, params=params, stream=True)

for chunk in response.iter_lines():
    if chunk:
        print(chunk.decode("utf-8"))
Asked By: Robert Ritz

||
Source

Answers:

  • In the ask_statesman function, change the yield current_response
    statement to yield {"data": current_response}. This will wrap each
    response line in a dictionary with a "data" key.
  • In the request_handler function, instead of returning the stream_response directly, return a generator expression that yields each response line from ask_statesman wrapped in a dictionary as shown above.
    Here’s the modified code:
import os
import time

import openai

import fastapi
from fastapi import Depends, HTTPException, status, Request
from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
from fastapi.responses import StreamingResponse

auth_scheme = HTTPBearer()
app = fastapi.FastAPI()

openai.api_key = os.environ["OPENAI_API_KEY"]


def ask_statesman(query: str):
    #prompt = router(query)
    
    completion_reason = None
    response = ""
    while not completion_reason or completion_reason == "length":
        openai_stream = openai.ChatCompletion.create(
            model="gpt-3.5-turbo",
            messages=[{"role": "user", "content": query}],
            temperature=0.0,
            stream=True,
        )
        for line in openai_stream:
            completion_reason = line["choices"][0]["finish_reason"]
            if "content" in line["choices"][0].delta:
                current_response = line["choices"][0].delta.content
                print(current_response)
                yield {"data": current_response}
                time.sleep(0.25)


@app.post("/")
async def request_handler(auth_key: str, query: str):
    if auth_key != "123":
        raise HTTPException(
            status_code=status.HTTP_401_UNAUTHORIZED,
            detail="Invalid authentication credentials",
            headers={"WWW-Authenticate": auth_scheme.scheme_name},
        )
    else:
        return StreamingResponse((line for line in ask_statesman(query)), media_type="text/plain")


if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000, debug=True, log_level="debug")
Answered By: Mahmoud alfayoumi

First, it wouldn’t be good practice to use a POST request for requesting data from the server. Using a GET request instead would be more suitable to your case. In addition to that, you shouldn’t be sending credentials, such as auth_key as part of the URL (i.e., using the query string), but you should rather use Headers and/or Cookies (using HTTPS). Please have a look at this answer for more details and examples on the concepts of headers and cookies, as well as the risks involved when using query parameters instead. Helpful posts around this topic can also be found here and here, as well as here, here and here.

Second, if you are executing a blocking operation (i.e., I/O-bound or CPU-bound tasks) inside the StreamingResponse‘s generator function, you should define the function with def (as you currently do) instead of async def, as, otherwise, the blocking operation, as well as the time.sleep() function that is used inside your generator, would blcok the event loop. As explained here, if the function for streaming the response body is a normal generator/iterator (i.e., def) and not an async def one, FastAPI will use iterate_in_threadpool() to run the iterator/generator in a separate thread that is then awaited. Have a look at this detailed answer as well for more details.

Third, you are using requests‘ iter_lines() function, which iterates over the response data, one line at a time. If, however, the response data do not include any line break character (i.e., n), you would not see the data on client’s console getting printed as they arrive, until the entire response is received by the client and printed as a whole. In that case, you should instead use iter_content() and specify the chunk_size as desired (both cases are demonstrated in the example below).

Finally, if you would like the StreamingResponse to work in every browser (including Chrome as well)—in the sense of being able to see the data as they stream in—you should specify the media_type to a different type than text/plain (e.g., text/event-stream instead), or disable MIME Sniffing. As explained here, browsers will start buffering text/plain responses for a certain amount (around 1445 bytes, as documented here), in order to check whether or not the content received is actually plain text. To avoid that, you can either set the media_type to text/event-stream (used for server-sent events), or keep using text/plain, but set the X-Content-Type-Options response header to nosniff, which would disable MIME Sniffing (both options are demonstrated in the example below).

Working Example

app.py

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
import time


app = FastAPI()


def fake_data_streamer():
    for i in range(10):
        yield b'some fake datann'
        time.sleep(0.5)


@app.get('/')
async def main():
    return StreamingResponse(fake_data_streamer(), media_type='text/event-stream')
    # or, use:
    #headers = {'X-Content-Type-Options': 'nosniff'}
    #return StreamingResponse(fake_data_streamer(), headers=headers, media_type='text/plain')

test.py (using Python requests)

import requests

url = "http://localhost:8000/"

with requests.get(url, stream=True) as r:
    for chunk in r.iter_content(1024):  # or, for line in r.iter_lines():
        if chunk:
            print(chunk)

test.py (using httpx—see this, as well as this and this for the benefits of using httpx over requests)

import httpx

url = 'http://127.0.0.1:8000/'
with httpx.stream('GET', url) as r:
    for chunk in r.iter_raw():  # or, for line in r.iter_lines():
        print(chunk)
Answered By: Chris

If you opt to use Langchain to interact with OpenAI (which I highly recommend), it provides stream method, which effectively returns a generator.

Slight modification to Chris‘ code above,

api.py

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from langchain.llms import OpenAI


llm = OpenAI(
    streaming=True,
    verbose=True,
    temperature=0,
)

app = FastAPI()


def chat_gpt_streamer(query: str):
    for resp in llm.stream(query):
        yield resp["choices"][0]["text"]


@app.get('/streaming/ask')
async def main(query: str):
    return StreamingResponse(chat_gpt_streamer(query), media_type='text/event-stream')

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000, log_level="debug")

Similarly you can test with httpx, or requests (again copy paste from Chris’ code):

test.py

import httpx

url = 'http://127.0.0.1:8000/streaming/ask?query=How are you, write in 10 sentences'
with httpx.stream('GET', url) as r:
    for chunk in r.iter_raw():  # or, for line in r.iter_lines():
        print(chunk)
Answered By: Mehmet Öner Yalçın

might consider looking into Server Send Events: https://github.com/sysid/sse-starlette

first install the library : pip install sse-starlette

from fastapi import FastAPI
from sse_starlette.sse import EventSourceResponse
import time


app = FastAPI()


def data_streamer():
    for i in range(10):
        yield f"_{i}_".encode("utf-8")
        time.sleep(1)


@app.get('/')
async def main():
    return EventSourceResponse(data_streamer(), media_type='text/event-stream')
Answered By: kiranr