Multi-translate is a unified interface on top of various translate APIs providing optimal translations ⭐, persistence 💾, fallback ♻️.

It is intended as a drop in replacement in any situation where you would integrate with an external translation API. With Multi-Translate you integrate once and get all the best of all the external translation APIs. Additionally you avoid expensive repeated requests with the built in persistence, and introduce redundancy with the ability to fallback to another service when a request fails.

Multi-Translate uses FastApi and asyncpg to keep things snappy ⚡, and offers graphql and json endpoints.

You can have a look and play at the website and functional demo here:

https://multi-translate.rekon.uk

You can check out our public instance's api docs here:

https://multi-translate-public-api.rekon.uk/docs

(redoc) https://multi-translate-public-api.rekon.uk/redoc

(graphql) https://multi-translate-public-api.rekon.uk/gql

Multi-translate aims to provide optimal translations for the given language pair. It does this by combining the strengths of the Amazon, Deep L, Google, Microsoft, Naver Papago, and Yandex translation APIs.

User's can specify which engine they want to use with the preferred_engine parameter. However there is also the default value of best for this field.

When best is used, the first suitable engines will be filtered out based on required capabilities - for instance if from_language is not specified and detection is required then engines which do not support detection will be eliminated. Similarly when with_alignment is set to true engines which don't support alignment will be eliminated. Finally, if the language pair is not supported for an engine it will be eliminated.

The controller will use then language_preferences.yaml (which is configurable) to look up the best translation engine for the from/to language combination based on the remaining languages. The default preferences are listed in default_preferred.md.

When a result is fetched for a particular engine, language, feature, and source text, it will be stored in a postgres database, if it has been fetched before, it will be retrieved from the database instead to avoid unnecessary usage charges. The write to the database takes place after the response is returned to keep responses fast ⚡.

The fallback option can be used so that if a result fails for the specified engine, for whatever reason, then the next best valid engine in the list will be chosen.

Rate limiting is optionally built in.

The helm chart is the easiest way to install

helm repo add rekon http://charts.rekon.uk
helm repo update
helm upgrade --install multi-translate rekon/multi-translate

Documentation for values can be found in charts/multi-translate/README.md

You can download the docker image with:

docker pull rekonuk/multi-translate

To determine which env variables to set you can look at settings.py and charts/multi-translate/README.md.

The core translate endpoint can be found at GET /translate or POST /translate. Documentation is available in swagger and redoc style at /docs and /redoc respectively.

import httpx
request_data = {
    "to_language": "es",
    "source_text": "How do you do?",
    "with_alignment": True,
}
resp = httpx.get("http://localhost:8080/translate", params=request_data)
post_resp = httpx.post("http://localhost:8080/translate", json=request_data)

assert resp.status_code == post_resp.status_code == 200
result = resp.json()
post_result = post_resp.json()

assert result == post_result == {
    "translated_text": "¿Cómo estás?",
    "engine": "microsoft",
    "engine_version": "3.0",
    "from_language": "en",
    "to_language": "es",
    "source_text": "hello",
    "detected_language_confidence": "1.0",
    "alignment": [
        {
            "dest": {"end": "4", "start": "0", "text": "안녕하세요"},
            "src": {"end": "4", "start": "0", "text": "hello"},
        }
    ],
}

The graphql endpoint is available (with GraphiQL UI) at /gql

query GetTranslation {
    translation(sourceText: "How do you do?", toLanguage: "es", withAlignment: true) {
        translatedText
        fromLanguage
        alignment{
          src{
            start,
            end,
            text
          }
          dest{
            start,
            end,
            text
          }
        }
  }
}

Result

{
  "data": {
    "translation": {
      "translatedText": "¿Cómo estás?",
      "fromLanguage": "en",
      "alignment": [
        {
          "src": {
            "start": 0,
            "end": 13,
            "text": "How do you do?"
          },
          "dest": {
            "start": 0,
            "end": 11,
            "text": "¿Cómo estás?"
          }
        }
      ]
    }
  }
}

Autogenerated OpenAPI 3.0 clients are available for go, python and javascript.

You can see examples and installation instructions here.

Run the app using skaffold

skaffold dev --port-forward

Run the functional tests once the app is running and port-forwarded

pytest functional_tests/ -vv

The helm chart documentation is autogenerated with helm-docs

docker run --rm -v "$(pwd):/helm-docs" -u $(id -u) jnorwood/helm-docs:latest

The api documentation is contained within the API Clients and will be hosted on your instance at /docs and /redoc.