LLM Streaming Responses Implementation – Turnkey

You launched an LLM-powered chatbot, and users are complaining about long response times. A 5–10 second delay kills conversion — each token appears almost instantly on the client, creating a live typing effect. This is critical for chatbots, assistants, and any interface where the user waits for a response. Once we implemented streaming for a fintech service with 10,000 concurrent users: after switching to streaming output, latency p99 dropped from 12 s to 400 ms, and conversion increased by 30%. Streaming output speeds up time to first response by 10–20 times compared to batch delivery. The solution is streaming responses. We implement LLM streaming using Server-Sent Events (SSE) and WebSocket, reducing visible latency to 200–500 ms. Our experience: 50+ LLM integrations for fintech, e-commerce, and SaaS. We guarantee quality: latency p99 < 500 ms.

According to MDN documentation, SSE uses a standard HTTP connection and automatically reconnects on drop. This makes it optimal for most chatbot scenarios.

How LLM Streaming Works

Instead of waiting for the full response, the server sends tokens to the client as they are generated. The protocol used is SSE (Server-Sent Events) — a text stream over HTTP for streaming data. The client receives data: chunks with JSON field text. When the model completes, it sends done: true. SSE is 2–3 times simpler to implement than WebSocket and requires no additional backend libraries. Without streaming, the user waits 5–10 seconds — long enough to leave. With streaming, the first character appears within 200 ms. This way, the user sees streaming output in real time.

Why SSE Over WebSocket?

SSE is simpler: no handshake, automatic reconnection, works over HTTP/2. WebSocket enables bidirectionality but is overkill for a simple chat. Comparison:

For 90% of cases, SSE is the optimal choice. WebSocket is justified for bidirectional transfer (e.g., voice commands).

Supported Models

Virtually all modern LLMs support streaming: Claude (Anthropic), GPT-4o (OpenAI), LLaMA 3 (via vLLM), Gemini (Google), Mistral, Qwen. For each model, the corresponding SDK with streaming mode must be used. For example, AsyncAnthropic and AsyncOpenAI provide async token generators. We choose the model for your use case: for chatbots — Claude Sonnet or GPT-4o, for real-time assistants — LLaMA 3 via vLLM for minimal latency.

Handling Interruptions and Backpressure

The client can cancel the request at any time. If not handled, the model continues generation — wasting tokens and money. We implement a cancel_event:

import asyncio

async def stream_with_cancellation(
    messages: list[dict],
    cancel_event: asyncio.Event,
) -> AsyncGenerator[str, None]:
    """Streaming with cancellation support"""
    async with anthropic_client.messages.stream(
        model="claude-sonnet-4-5",
        max_tokens=2048,
        messages=messages,
    ) as stream:
        async for text in stream.text_stream:
            if cancel_event.is_set():
                stream.close()
                yield f"data: {json.dumps({'cancelled': True})}\n\n"
                return
            yield f"data: {json.dumps({'text': text})}\n\n"

Additionally, we use backpressure: when the client buffer is full, we pause the generator iteration. This prevents connection drops.

Backend: FastAPI + SSE

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from anthropic import AsyncAnthropic
from openai import AsyncOpenAI
import asyncio
import json

app = FastAPI()
anthropic_client = AsyncAnthropic()
openai_client = AsyncOpenAI()

async def stream_anthropic(messages: list[dict], system: str = "") -> AsyncGenerator[str, None]:
    """Generator for Claude streaming"""
    async with anthropic_client.messages.stream(
        model="claude-sonnet-4-5",
        max_tokens=2048,
        system=system,
        messages=messages,
    ) as stream:
        async for text in stream.text_stream:
            yield f"data: {json.dumps({'text': text})}\n\n"
        yield f"data: {json.dumps({'done': True})}\n\n"

async def stream_openai(messages: list[dict]) -> AsyncGenerator[str, None]:
    """Generator for OpenAI streaming"""
    async with await openai_client.chat.completions.create(
        model="gpt-4o",
        messages=messages,
        stream=True,
    ) as stream:
        async for chunk in stream:
            delta = chunk.choices[0].delta
            if delta.content:
                yield f"data: {json.dumps({'text': delta.content})}\n\n"
        yield f"data: {json.dumps({'done': True})}\n\n"

@app.post("/chat/stream")
async def chat_stream(request: dict):
    messages = request.get("messages", [])
    provider = request.get("provider", "anthropic")

    generator = (
        stream_anthropic(messages)
        if provider == "anthropic"
        else stream_openai(messages)
    )

    return StreamingResponse(
        generator,
        media_type="text/event-stream",
        headers={
            "Cache-Control": "no-cache",
            "X-Accel-Buffering": "no",
        }
    )

We choose AsyncAnthropic and AsyncOpenAI for async streaming. Important: set X-Accel-Buffering: no for nginx — otherwise nginx will buffer SSE.

Frontend: React with Streaming Read

import { useState, useCallback } from 'react';

function useStreamingChat() {
  const [response, setResponse] = useState('');
  const [isStreaming, setIsStreaming] = useState(false);

  const sendMessage = useCallback(async (messages: Message[]) => {
    setIsStreaming(true);
    setResponse('');

    const res = await fetch('/chat/stream', {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({ messages }),
    });

    const reader = res.body!.getReader();
    const decoder = new TextDecoder();

    while (true) {
      const { done, value } = await reader.read();
      if (done) break;

      const chunk = decoder.decode(value);
      const lines = chunk.split('\n\n');

      for (const line of lines) {
        if (line.startsWith('data: ')) {
          const data = JSON.parse(line.slice(6));
          if (data.done) {
            setIsStreaming(false);
            return;
          }
          setResponse(prev => prev + data.text);
        }
      }
    }
    setIsStreaming(false);
  }, []);

  return { response, isStreaming, sendMessage };
}

For POST requests we use ReadableStream — more flexible than EventSource, and allows sending a request body.

Process

1. Analytics: review your architecture, select model and protocol (SSE/WebSocket).
2. Design: draw flow diagram, define interruption points.
3. Implementation: write backend endpoint, frontend component, error handling.
4. Testing: load test with simulated N users, measure p99 latency.
5. Deployment & monitoring: configure nginx, CDN, monitoring (Grafana + Prometheus) to track streaming connections.

What's Included

• Backend endpoint with SSE or WebSocket (FastAPI).
• React component (TypeScript) with interruption handling, loading indicator, auto-scroll.
• Documentation: protocol description, nginx configs, request examples.
• Post-launch support: 1 month incident management.

Estimated Timelines

Cost is calculated individually. Contact us for a project estimate. Order streaming output implementation for your LLM service — improve UX and conversion.

Streaming output is not just a feature, but a requirement for modern UX of streaming chatbots. We implement it turnkey in 5–10 days. Our experience: 50+ LLM projects. Get a consultation right now.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.