Email Classifier Pipeline with LLM and IMAP

You will build a pipeline that connects to your email server via IMAP, classifies incoming messages by urgency and department, extracts key information, and generates draft responses — all using a self-hosted LLM. The end result: support@yourcompany.com receives 200 emails daily, and each is automatically classified (billing, technical, complaint, spam), prioritised (urgent, normal, low), and routed to the right team with a suggested response draft. No email content leaves your server. Here is the complete pipeline on dedicated GPU infrastructure.

Pipeline Architecture

Email Fetching

import imaplib
import email
from email.header import decode_header

def fetch_unprocessed_emails(server, username, password, folder="INBOX"):
    mail = imaplib.IMAP4_SSL(server)
    mail.login(username, password)
    mail.select(folder)

    _, messages = mail.search(None, "UNFLAGGED")
    emails = []
    for msg_id in messages[0].split()[-50:]:  # Process 50 at a time
        _, data = mail.fetch(msg_id, "(RFC822)")
        msg = email.message_from_bytes(data[0][1])
        body = extract_body(msg)
        emails.append({
            "id": msg_id,
            "from": msg["From"],
            "subject": decode_header(msg["Subject"])[0][0],
            "body": body[:2000],  # Truncate for LLM context
            "date": msg["Date"]
        })
    return emails

LLM Classification

from openai import OpenAI
import json

client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

def classify_email(email_data: dict) -> dict:
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.1-8B-Instruct",
        messages=[{
            "role": "system",
            "content": """Classify this email. Return JSON:
{"category": "billing|technical|complaint|sales|spam|other",
 "urgency": "urgent|normal|low",
 "sentiment": "positive|negative|neutral",
 "key_entities": {"customer_name": "", "account_id": "", "product": ""},
 "summary": "One sentence summary",
 "suggested_team": "billing|engineering|customer_success|sales"}"""
        }, {
            "role": "user",
            "content": f"From: {email_data['from']}\n"
                       f"Subject: {email_data['subject']}\n"
                       f"Body: {email_data['body']}"
        }],
        max_tokens=300, temperature=0.0
    )
    return json.loads(response.choices[0].message.content)

The vLLM server handles batch classification efficiently. Temperature 0 ensures consistent categorisation across similar emails.

Draft Response Generation

def generate_draft(email_data: dict, classification: dict) -> str:
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.1-8B-Instruct",
        messages=[{
            "role": "system",
            "content": f"Draft a professional response to this {classification['category']} email. "
                       f"Be empathetic if the sentiment is negative. "
                       f"Include relevant next steps. Keep under 150 words. "
                       f"Sign as 'Customer Support Team'."
        }, {
            "role": "user",
            "content": f"Subject: {email_data['subject']}\nBody: {email_data['body']}"
        }],
        max_tokens=300, temperature=0.3
    )
    return response.choices[0].message.content

Routing and Automation

Route classified emails based on urgency and category. Urgent complaints go directly to a senior support queue with the draft response attached. Billing queries route to the finance team with extracted account details. Spam gets flagged and archived. Normal queries enter the standard support queue with the AI draft pre-loaded for agent editing. Mark processed emails with an IMAP flag so they are not reprocessed.

Production Deployment

For production: run classification every 2 minutes via a cron job; implement a confidence threshold — if the LLM is uncertain, route to a human triage queue; log all classifications for accuracy monitoring; never auto-send draft responses without human approval; and add RAG retrieval from your knowledge base to improve response quality. Teams handling sensitive correspondence should deploy on private infrastructure with encryption. See model options for multilingual email handling, chatbot hosting for real-time customer communication, more tutorials, and industry use cases. Check infrastructure guides for IMAP security.