How to Build an AI Knowledge Article Suggester

An ai knowledge article suggestion system for support agents surfaces the right help article the moment a ticket opens. I build these to cut the time agents spend searching through documentation. Instead of digging through a knowledge base, the system reads the ticket and suggests the 3 most relevant articles with confidence scores.

Agents who use suggested articles resolve tickets 40% faster. That is the difference between searching and finding.

What You Need Before Starting

• A knowledge base with articles in a searchable format
• Python 3.8+ with sentence-transformers or an embedding API
• A vector store (ChromaDB for local, Pinecone for production)
• Your ticketing system's webhook or API

Step 1: Index Your Knowledge Base

Convert articles to embeddings for semantic search:

import chromadb
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")
chroma = chromadb.PersistentClient(path="./kb_vectors")
collection = chroma.get_or_create_collection("knowledge_base")

def index_articles(articles):
    for article in articles:
        text = f"{article['title']} {article['content']}"
        embedding = model.encode(text).tolist()
        collection.add(
            ids=[article["id"]],
            embeddings=[embedding],
            documents=[text],
            metadatas=[{"title": article["title"], "url": article["url"], "category": article["category"]}]
        )

Step 2: Build the Suggestion Engine

When a ticket comes in, find the closest matching articles:

def suggest_articles(ticket_subject, ticket_body, top_k=3):
    query_text = f"{ticket_subject} {ticket_body}"
    query_embedding = model.encode(query_text).tolist()

    results = collection.query(
        query_embeddings=[query_embedding],
        n_results=top_k
    )

    suggestions = []
    for i in range(len(results["ids"][0])):
        suggestions.append({
            "id": results["ids"][0][i],
            "title": results["metadatas"][0][i]["title"],
            "url": results["metadatas"][0][i]["url"],
            "relevance_score": round(1 - results["distances"][0][i], 3)
        })

    return suggestions

Step 3: Hook into the Ticket Workflow

Trigger suggestions automatically on new tickets:

from flask import Flask, request, jsonify

app = Flask(__name__)

@app.route("/ticket-created", methods=["POST"])
def on_ticket_created():
    ticket = request.json

    suggestions = suggest_articles(ticket["subject"], ticket["body"])

    add_internal_note(ticket["id"], format_suggestions(suggestions))

    return jsonify({"ticket_id": ticket["id"], "suggestions": suggestions})

def format_suggestions(suggestions):
    lines = ["Suggested knowledge articles:"]
    for s in suggestions:
        lines.append(f"- [{s['title']}]({s['url']}) (relevance: {s['relevance_score']})")
    return "\n".join(lines)

Step 4: Track Which Suggestions Get Used

Log when agents click a suggested article to measure usefulness:

import sqlite3

def log_suggestion_used(ticket_id, article_id):
    conn = sqlite3.connect("suggestions.db")
    conn.execute("""
        INSERT INTO suggestion_usage (ticket_id, article_id, used_at)
        VALUES (?, ?, datetime('now'))
    """, (ticket_id, article_id))
    conn.commit()

def get_usage_stats():
    conn = sqlite3.connect("suggestions.db")
    total_suggested = conn.execute("SELECT COUNT(*) FROM suggestions").fetchone()[0]
    total_used = conn.execute("SELECT COUNT(*) FROM suggestion_usage").fetchone()[0]
    return {"usage_rate": round(total_used / total_suggested * 100, 1) if total_suggested else 0}

Step 5: Re-Index on Content Updates

Keep suggestions fresh by re-indexing when articles change:

def update_article_index(article):
    text = f"{article['title']} {article['content']}"
    embedding = model.encode(text).tolist()
    collection.update(
        ids=[article["id"]],
        embeddings=[embedding],
        documents=[text],
        metadatas=[{"title": article["title"], "url": article["url"], "category": article["category"]}]
    )

What to Build Next

Add a feedback button. When an agent marks a suggestion as "not helpful," log that data and use it to fine-tune your embedding model or add missing articles to the knowledge base.

Related Reading

• The Centralized Brain Concept - your knowledge base as a centralized intelligence layer
• AI in Customer Service - knowledge suggestions in the support automation stack
• How Systems Compound Over Time - better suggestions compound into faster resolutions