How to detect duplicate support tickets with embeddings

Overview

Fikra API is an OpenAI-compatible inference API. You send ticket text to the embeddings endpoint, get one vector per ticket, and compare vectors to surface near-duplicates before an agent picks them up. The same base URL and key also serve chat models, so you can extend the flow later.

Prerequisites

• A Fikra API key. No credit-card gate to get started; top up the token-credit wallet when you are ready.
• The key exported as FIKRA_API_KEY in your environment. Keep it server-side, never in client code.
• Python with the OpenAI SDK installed (pip install openai).
• A batch of support ticket texts to embed.

Step-by-step

1. Export your key and create a client pointed at the Fikra base URL.
2. Send the ticket texts to POST https://api.fikraapi.co.ke/v1/embeddings in one request.
3. Compute cosine similarity between every pair of vectors.
4. Flag pairs above a threshold you tune on your own data.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.fikraapi.co.ke/v1",
    api_key=os.environ["FIKRA_API_KEY"],
)

tickets = [
    "I was charged twice for my subscription this month.",
    "Duplicate charge on my card for the monthly plan.",
    "How do I reset my password?",
]

response = client.embeddings.create(
    model="fikra-auto",
    input=tickets,
)

vectors = [item.embedding for item in response.data]

Model availability and per-model pricing are on the pricing page , confirm the model you pass before production use.

import math

def cosine(a, b):
    dot = sum(x * y for x, y in zip(a, b))
    na = math.sqrt(sum(x * x for x in a))
    nb = math.sqrt(sum(y * y for y in b))
    return dot / (na * nb)

THRESHOLD = 0.90  # starting point only; tune against your own tickets

for i in range(len(vectors)):
    for j in range(i + 1, len(vectors)):
        score = cosine(vectors[i], vectors[j])
        if score >= THRESHOLD:
            print(f"possible duplicate: {i} <-> {j} score={score:.3f}")

Requests and responses follow the OpenAI-style shape, so existing OpenAI SDK code works once you change base_url and the key.

Verify it works

• len(response.data) equals the number of tickets you sent, and every embedding has the same length.
• Each item carries its index, so you can map vectors back to tickets.
• With the sample above, the two billing tickets should be flagged and the password ticket should not.
• If a call fails, check the status code against the list below before changing your code.

Costs

Usage is billed per token against your token-credit wallet, in USD or KES. Pricing starts at $0.09 per 1M input tokens and $0.15 per 1M output tokens. Listed examples:

Top up by M-Pesa or card through Paystack, or with x402 USDC on Base for agents. Rate limits are 30 requests/min for unverified accounts and 100 requests/min for trusted accounts. For anything not listed here, check the pricing page.

Common errors

• 401 Unauthorized , the key is missing, wrong, or inactive. Confirm FIKRA_API_KEY holds a current key; keys are scoped and revocable.
• 400 Bad Request , unsupported or misspelled model. Use a model name from the pricing page.
• 429 Too Many Requests , you hit the rate limit (30 requests/min unverified, 100 requests/min trusted). Batch tickets into fewer embedding calls.
• 503 Service Unavailable , upstream is busy. Retry with exponential backoff.

Last updated 2026-10-18.