Skip to content
OLAVERSELABDocs

Route messages by language

Know what language a message is in before you act on it, and route it to someone, or something, that can read it.

General-purpose detectors mislabel short African-language messages: a Yorùbá greeting comes back as English, Hausa as “unknown”. Language identification covers 608 languages, and its default traffic mode is tuned for short, messy user input. This script checks a batch of messages in one call and assigns each a queue.

Run it

Section titled “Run it”
Terminal window
pip install requests
export OLAVERSE_KEY="your-key"     # from https://console.olaverselab.com
python route.py

The script

Section titled “The script”
route.py
import os


import requests


URL = os.environ.get("OLAVERSE_URL", "https://api.olaverselab.com")
HEADERS = {"Authorization": f"Bearer {os.environ['OLAVERSE_KEY']}"}




def call(path, body):
    r = requests.post(f"{URL}{path}", json=body, headers=HEADERS, timeout=120)
    r.raise_for_status()
    return r.json()




# ISO 639-3 codes, as the API returns them, to your queues.
QUEUES = {"yor": "yoruba-agents", "hau": "hausa-agents", "ibo": "igbo-agents", "eng": "english-agents"}
SURE = 0.6   # below this the text is usually too short to tell ("ok", "hi")




def route(messages):
    # One batch call for up to 256 messages; results come back in the same order.
    results = call("/v1/language-id/batch", {"texts": messages})["results"]
    routed = []
    for message, r in zip(messages, results):
        if r["language"] == "zxx" or r["confidence"] < SURE:
            queue = "general"                      # code, numbers, or too short
        else:
            queue = QUEUES.get(r["language"], "general")
        routed.append((queue, r["language"], round(r["confidence"], 2), message))
    return routed




if __name__ == "__main__":
    messages = [
        "Ina son in canza lambar wayata",
        "Mo fẹ́ mọ ìgbà tí ẹrù mi máa dé",
        "Biko, enyemaka m chọrọ maka akaụntụ m",
        "Please cancel my order",
        "12345",
    ]
    for queue, language, confidence, message in route(messages):
        print(f"{queue:15} {language:4} {confidence:5}  {message}")

Response

Section titled “Response”

What /v1/language-id/batch returns for each message, in order. Values are illustrative.

{"results": [
  {"language": "hau", "script": "Latn", "label": "hau_Latn", "language_name": "Hausa", "confidence": 0.98,
   "top": [{"label": "hau_Latn", "language": "hau", "language_name": "Hausa", "score": 0.98}], "warnings": [],
   "model": "lid-neural-608", "mode": "traffic", "usage": {"texts": 1}},
  ...
]}

The script prints one line per message, for example hausa-agents hau 0.98 Ina son in canza lambar wayata.

How the answer is used

Section titled “How the answer is used”
  • language is ISO 639-3 (yor, hau, ibo, swh); label adds the script (yor_Latn).
  • zxx means no language: the text is code, numbers or symbols, like 12345.
  • Low confidence on very short text (“ok”, “hi”) is expected; below SURE the script sends it to general. Route those by the conversation’s earlier messages if you have them.

Adapt it

Section titled “Adapt it”
  • Then decide: add the detected language to the state of a decision, or use it to pick the language of your reply.
  • Restore tone marks for display: messages typed without diacritics read better after POST /v1/diacritize with the detected language (Yorùbá, Igbo, Hausa and 8 others).
  • Fastest and cheapest: add "model": "lid-lite-608". For corpora where every language should count equally, "mode": "coverage".
  • Long messages: each text can be up to 1,000 characters; the first 1,000 are plenty to tell the language.
Language identificationModels, modes and fields.
Language routing use caseThe full pipeline and benchmark figures.