Multilingual vs General-Purpose LLM Models: Understanding the Differences

#learnai #oxlo #ai

I built a multilingual support router that classifies incoming tickets by language and routes them to either a general-purpose or a multilingual specialist model. This helps teams with mixed-language queues reduce costs on long tickets by relying on flat per-request inference instead of token-based routing layers. The whole pipeline runs against Oxlo.ai's OpenAI-compatible endpoint.

What you'll need

Python 3.10 or newer
An Oxlo.ai API key from https://portal.oxlo.ai
The OpenAI SDK installed with pip install openai

Step 1: Initialize the Oxlo.ai client

I start by importing the OpenAI SDK and pointing it at Oxlo.ai. Because the platform is fully OpenAI SDK compatible, this is a drop-in replacement and no extra adapters are needed.

from openai import OpenAI
import json

client = OpenAI(base_url="https://api.oxlo.ai/v1", api_key="YOUR_OXLO_API_KEY")

Step 2: Define the agent's system prompt

This prompt is shared by both model paths. It instructs the agent to match the user's language and tone, and to keep responses concise. I keep it single-shot so it works unchanged regardless of which model picks it up.

SYSTEM_PROMPT = """You are a senior SaaS support agent.
Read the ticket below and reply in the exact language the user wrote in.
If the user wrote in English, give a direct, technical answer.
If the user wrote in another language, preserve local technical terminology and polite formality.
Keep your answer under three sentences."""

Step 3: Build the triage function with Qwen 3 32B

I use Qwen 3 32B for the triage step because it handles multilingual reasoning and agent workflows well. The model returns strict JSON so the rest of the script can route deterministically, and I enable JSON mode to avoid parsing drift.

TRIAGE_PROMPT = """You are a routing layer. Read the support ticket and return only a JSON object with this exact schema:
{"language": "ISO-639-1 code", "topic": "one-word category", "urgency": "low|medium|high"}
Do not add markdown or explanation."""

def triage_ticket(text: str) -> dict:
    response = client.chat.completions.create(
        model="qwen-3-32b",
        messages=[
            {"role": "system", "content": TRIAGE_PROMPT},
            {"role": "user", "content": text},
        ],
        response_format={"type": "json_object"},
    )
    raw = response.choices[0].message.content
    return json.loads(raw)

Step 4: Route to the right model and generate the answer

If the ticket is English, I send it to Llama 3.3 70B for fast, general-purpose inference. For any other language, I keep it on Qwen 3 32B. Because Oxlo.ai charges per request rather than per token, splitting the workload across two calls does not scale costs with ticket length. You can see the pricing details at https://oxlo.ai/pricing.

def generate_response(ticket_text: str, triage: dict) -> tuple[str, str]:
    model = "llama-3.3-70b" if triage["language"] == "en" else "qwen-3-32b"
    response = client.chat.completions.create(
        model=model,
        messages=[
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": f"Language: {triage['language']}\nTopic: {triage['topic']}\nUrgency: {triage['urgency']}\nTicket: {ticket_text}"},
        ],
    )
    return model, response.choices[0].message.content

Run it

I test the pipeline on three tickets: English, Spanish, and Japanese. The script prints which model handled each one so you can see the routing in action.

if __name__ == "__main__":
    tickets = [
        "My API key started returning 401 errors immediately after rotation.",
        "El dashboard no carga después de la actualización de ayer, necesito ayuda urgente.",
        "昨日からレポートのエクスポートが失敗します。至急確認してください。",
    ]

    for t in tickets:
        triage = triage_ticket(t)
        model, answer = generate_response(t, triage)
        print(f"Ticket: {t[:45]}...")
        print(f"Triage: {triage}")
        print(f"Routed to: {model}")
        print(f"Response: {answer}\n")

Example output:

Ticket: My API key started returning 401 errors immedia...
Triage: {'language': 'en', 'topic': 'authentication', 'urgency': 'high'}
Routed to: llama-3.3-70b
Response: Check that your new key is active in the Oxlo.ai portal and that you are using the correct base URL, https://api.oxlo.ai/v1. Clear any cached keys and retry.

Ticket: El dashboard no carga después de la actualizaci...
Triage: {'language': 'es', 'topic': 'dashboard', 'urgency': 'high'}
Routed to: qwen-3-32b
Response: Entiendo la urgencia. Por favor, intenta limpiar la caché del navegador y verifica el estado del servicio en nuestra página de estado. Si el problema persiste, envíanos los logs de la consola.

Ticket: 昨日からレポートのエクスポートが失敗します。至急確認してください。...
Triage: {'language': 'ja', 'topic': 'export', 'urgency': 'high'}
Routed to: qwen-3-32b
Response: ご連絡ありがとうございます。エクスポート設定とファイル形式を確認いたします。しばらくお待ちいただくか、サポートチャットでトランザクションIDをお送りください。

Wrap-up

You now have a working router that treats multilingual tickets as first-class workloads instead of forcing them through a general-purpose pipeline. For a production version, cache the triage step with a lightweight embedding check using Oxlo.ai's BGE-Large endpoint, or swap in DeepSeek V3.2 for English code tickets and Kimi K2.6 for vision-heavy agentic workflows.

DEV Community