LLM streaming FASTAPI

main.py

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from fastapi.staticfiles import StaticFiles  
from langchain.callbacks.streaming_aiter import AsyncIteratorCallbackHandler
from langchain.callbacks.manager import AsyncCallbackManager  
from langchain.chat_models import init_chat_model
import asyncio
import os

GROQ_API_KEY = "key"
os.environ["GROQ_API_KEY"] = GROQ_API_KEY

app = FastAPI()

app.mount("/static", StaticFiles(directory=".", html=True), name="static")


@app.get("/stream")
async def stream_response(query: str):
    callback = AsyncIteratorCallbackHandler()
    manager = AsyncCallbackManager([callback])

    llm = init_chat_model(
        model="groq:llama-3.1-8b-instant",
        streaming=True,
        callback_manager=manager,
        temperature=0
    )

    async def sse_generator():
        task = asyncio.create_task(llm.apredict(query))
        async for token in callback.aiter():
            yield f"data: {token}\n\n"
        yield "data: [DONE]\n\n"
        await task

    return StreamingResponse(sse_generator(), media_type="text/event-stream")

index.html

<!DOCTYPE html>
<html lang="en">

<head>
    <meta charset="UTF-8">
    <title>LLM Streaming Demo</title>
</head>

<body>
    <h1>LLM Streaming Demo</h1>
    <div id="output"></div>
    <script>
        const outputDiv = document.getElementById("output");
        const eventSource = new EventSource("/stream?query=Tell me a story about a brave knight");
        eventSource.onmessage = function (event) {
            console.log("Token:", event.data);
            if (event.data === "[DONE]") {
                eventSource.close();
            } else {
                outputDiv.textContent += event.data;
            }
        };
    </script>
</body>

</html>

SSE Format: What the Browser Expects

Per the SSE spec (WHATWG + HTML Living Standard), each event must follow this format:

data: message line 1
data: message line 2
event: optional-event-type
id: optional-id

To break it down:

Each event must end with two newlines (\n\n) — this signals the browser that the message is complete.
The data: prefix is required for each line of the message.
The browser will call EventSource.onmessage with the combined lines of data: fields, joined with \n.

Streaming Gotchas in GenAI Apps

1. Text-only protocols (SSE)

Gotcha: SSE only supports UTF-8 text — no binary.
Impact: You must serialize JSON and can't send files, images, or audio directly.
Tip: Use json.dumps() carefully and keep lines intact.

2. Partial message boundaries

Gotcha: SSE messages must end with \n\n. If you forget this, the frontend may never receive the message.
Impact: Frontend appears frozen even though server is sending data.
Tip: Always end each chunk with \n\n. In Python: yield f"data: {chunk}\n\n"

3. Buffering delays

Gotcha: Server or proxy buffering can delay delivery (Nginx, Gunicorn, Cloudflare).
Impact: Tokens generated are stuck in buffer and not flushed.
Fixes:
- Use flush=True (e.g., sys.stdout.flush() or response.flush()).
- Disable buffering at proxy level (e.g., proxy_buffering off; in Nginx).
- Consider uvicorn --http h11 for lower-level streaming in dev.

4. Browser limits and behavior

Gotcha: SSE can silently reconnect, replay events, or break if messages are malformed.
Impact: You may accidentally duplicate or drop tokens.
Fixes:
- Use id: in SSE messages for replay safety.
- Catch and deduplicate based on event.lastEventId if needed.

5. Error handling during generation

Gotcha: If your LLM toolchain fails midway (e.g., OpenAI rate limit or tool error), stream dies without clear signal.
Fix: Send custom event: error blocks or a [ERROR] marker with SSE.
Frontend Tip: Implement eventSource.onerror or a timeout.

6. Token join glitches

Gotcha: Tokens may be split mid-word or mid-sentence, especially if using a tokenizer.
Impact: Client sees broken words like "hel" "lo" or "Th" "is".
Fixes:
- Join tokens before display (+=), but clean up later.
- Optionally add a debounce or small buffer to group tokens.

7. LLM context truncation or restart

Gotcha: If the model restarts due to context overflow, the stream might appear to “repeat” earlier responses.
Tip: Always track and trim the token history sent to the LLM (e.g., sliding window).

8. Streaming across agent steps / tool calls

Gotcha: LLM stops mid-stream to call a tool, then resumes — you must handle this explicitly.
Fix:
- Use LangChain's AsyncIteratorCallbackHandler.
- Manually yield tool responses if needed.

9. Async / concurrency pitfalls

Gotcha: Blocking I/O (e.g., slow tool calls or DB fetches) will stall the stream.
Tip: Use async def, await, and asyncio.create_task() properly in your generators.

10. Frontend rendering latency

Gotcha: DOM updates for every token cause slow UI or flicker.
Fixes:
- Use requestAnimationFrame or throttled rendering.
- Buffer tokens and flush to DOM in small batches (e.g., 50ms intervals).

Bonus: Security + Compliance

Risk

Fix

SSE may leak tokens in error messages

Sanitize all logs and tool responses

Overstreaming can DDoS your client

Add rate limits or chunk caps

Don’t stream sensitive info token-by-token

Filter or post-process before sending

TL;DR Best Practices

Area

Practice

Backend

Use StreamingResponse, flush properly, break data cleanly

Frontend

Use EventSource, onmessage, debounce rendering

Error Handling

Send [DONE] and [ERROR], reconnect if needed

Start fast (first token in <500ms), stream naturally

Tool Calls

Interleave tool results with generation steps

PreviousAI Engg

Last updated 3 months ago

Was this helpful?