llmtranslate

llmtranslate

The llmtranslate library is a Python tool that leverages LangChain's ChatModels 🦜🔗 to translate text between languages and recognize the language of a given text. This library provides both synchronous and asynchronous translators to accommodate various use cases.

📖 Documentation: llm-translate.com

Installation

To ensure proper dependency management, install the required libraries in the following order:

• Install LangChain chat libraries for your preferred LLM. For example:

  pip install langchain-openai
  

  For other LLM integrations, refer to LangChain ChatModels documentation.

• Install the llmtranslate library:

  pip install llmtranslate
  

Poetry problem

If you’re using Poetry, you must set a Python requirement capped below 4.0 to match langchain-openai (which requires <4.0). So, if your pyproject.toml has:

python = ">=3.12"

replace it with:

python = "^3.12"

or:

python = ">=3.12,<4.0"

Otherwise, Poetry sees Python 4.x as valid, which conflicts with langchain-openai’s <4.0 requirement.

Example Using OpenAI's GPT

Using Translator (Synchronous)

from llmtranslate import Translator
from langchain_openai import ChatOpenAI

# Initialize the LLM and Translator
llm = ChatOpenAI(model_name="gpt-4o", openai_api_key="your_openai_api_key")
translator = Translator(llm=llm)

# Detect the language of the text
text_language = translator.get_text_language("Hi how are you?")
if text_language:
    print(text_language.ISO_639_1_code) # Output: en
    print(text_language.ISO_639_2_code) # Output: eng
    print(text_language.ISO_639_3_code) # Output: eng
    print(text_language.language_name) # Output: English

# Translate text
translated_text = translator.translate("Bonjour tout le monde", "English")
print(translated_text)  # Output: Hello everyone

Using AsyncTranslator (Asynchronous)

This example demonstrates how to use the AsyncTranslator class to asynchronously detect the language of a text and translate it into another language. The AsyncTranslator is designed to work with asynchronous operations by splitting the input text into manageable chunks and by handling multiple concurrent requests to your language model.

Key Configuration Options

• max_translation_chunk_length: Sets the maximum length for each text chunk that is translated.
• max_translation_chunk_length_multilang: Sets the maximum length for text chunks when dealing with multiple languages.
• max_concurrent_llm_calls: Limits the number of concurrent calls made to the underlying language model.
• max_concurrent_time_period: (New) Limits the maximum number of concurrent LLM calls within a specific time period (in seconds). For instance, if you want to allow up to 50 requests per minute, set:
  • max_concurrent_llm_calls to 50
  • max_concurrent_time_period to 60

If you only need to limit the number of concurrent calls without a time constraint, you can leave max_concurrent_time_period at its default value (typically 0).

Code Example

import asyncio
from llmtranslate import AsyncTranslator
from langchain_openai import ChatOpenAI

# Initialize the language model (LLM) with your API key
llm = ChatOpenAI(model_name="gpt-4o", openai_api_key="your_openai_api_key")

async def translate_text():
    # Initialize AsyncTranslator with appropriate parameters
    translator = AsyncTranslator(
        llm=llm,
        max_translation_chunk_length=100,           # Maximum chunk length for translation
        max_translation_chunk_length_multilang=50,    # Maximum chunk length for multi-language texts
        max_concurrent_llm_calls=10,                  # Maximum number of concurrent LLM calls
        # To enforce a rate limit (e.g., 50 calls per minute), uncomment and adjust the following:
        # max_concurrent_llm_calls=50,
        # max_concurrent_time_period=60,
    )
    
    # Prepare tasks: one to detect language and one to translate the text
    tasks = [
        translator.get_text_language("Hi how are you?"),
        translator.translate("Hi how are you?", "Spanish")
    ]
    
    # Run the tasks concurrently
    results = await asyncio.gather(*tasks)
    
    # Retrieve and display detected language information
    text_language = results[0]
    if text_language:
        print(text_language.ISO_639_1_code)  # Expected Output: en
        print(text_language.ISO_639_2_code)  # Expected Output: eng
        print(text_language.ISO_639_3_code)  # Expected Output: eng
        print(text_language.language_name)   # Expected Output: English

    # Display the translated text
    print(results[1])  # Expected Output: Hola, ¿cómo estás?

# Run the asynchronous translation process
asyncio.run(translate_text())

Summary

• Asynchronous Execution: Uses asyncio to concurrently run language detection and translation tasks.
• Language Detection: The get_text_language function returns language details (e.g., ISO 639-1, ISO 639-2, ISO 639-3 codes and language name).
• Translation: The translate function converts the input text to the desired target language.
• Concurrency Control:
  • Without Time Limit: Simply set max_concurrent_llm_calls to limit concurrent calls.
  • With Time Limit: Specify both max_concurrent_llm_calls and max_concurrent_time_period to control the number of calls within a specific time window (e.g., 50 calls per 60 seconds).

This example should help you understand how to integrate and configure all available parameters in the AsyncTranslator, including the new time-based rate limiting feature.

Key Parameters

max_translation_chunk_length

• Description: Defines the maximum length (in characters) of a text chunk to be translated in a single call when the text is in one language.
• Recommendation: If translations are not accurate, try reducing this value as weaker LLMs struggle with large chunks of text.

max_translation_chunk_length_multilang

• Description: Defines the maximum length (in characters) of a text chunk when the text contains multiple languages.
• Recommendation: Reduce this value for better accuracy with multi-language inputs.

max_concurrent_llm_calls

• Description: Limits the number of concurrent calls made to the LLM.
• Default: 10
• Usage: Adjust this parameter to align with your LLM provider's concurrency limits.

Other Supported LLMs Examples

You can find more examples of LangChain Chat models in the documentation at:

• llm-translate.com
• LangChain Documentation

Anthropic's Claude

To create an instance of an Anthropic-based Chat LLM:

  pip install langchain-anthropic
  pip install llmtranslate

from langchain_anthropic import Anthropic
from llmtranslate import Translator
llm = Anthropic(model="claude-2", anthropic_api_key="your_anthropic_api_key")
translator = Translator(llm=llm)

Mistral via API

To create an instance of a Chat LLM using the Mistral API:

pip install -qU langchain_mistralai
pip install llmtranslate

from langchain_mistral import MistralChat
from llmtranslate import Translator

llm = MistralChat(api_key="your_mistral_api_key")
translator = Translator(llm=llm)

Supported Languages

llmtranslate supports all languages supported by GPT-4o. For a complete list of language codes, you can visit the ISO 639-1 website.

Here is a table showing which languages are supported by gpt-4o and gpt4o-mini:

Language Name	Language Code	Supported by gpt-4o	Supported by gpt4o-mini
English	en	Yes	Yes
Mandarin Chinese	zh	Yes	Yes
Hindi	hi	Yes	Yes
Spanish	es	Yes	Yes
French	fr	Yes	Yes
German	de	Yes	Yes
Russian	ru	Yes	Yes
Arabic	ar	Yes	Yes
Italian	it	Yes	Yes
Korean	ko	Yes	Yes
Punjabi	pa	Yes	Yes
Bengali	bn	Yes	Yes
Portuguese	pt	Yes	Yes
Indonesian	id	Yes	Yes
Urdu	ur	Yes	Yes
Persian (Farsi)	fa	Yes	Yes
Vietnamese	vi	Yes	Yes
Polish	pl	Yes	Yes
Samoan	sm	Yes	Yes
Thai	th	Yes	Yes
Ukrainian	uk	Yes	Yes
Turkish	tr	Yes	Yes
Maori	mi	No	No
Norwegian	no	Yes	Yes
Dutch	nl	Yes	Yes
Greek	el	Yes	Yes
Romanian	ro	Yes	Yes
Swahili	sw	Yes	Yes
Hungarian	hu	Yes	Yes
Hebrew	he	Yes	Yes
Swedish	sv	Yes	Yes
Czech	cs	Yes	Yes
Finnish	fi	Yes	Yes
Amharic	am	No	No
Tagalog	tl	Yes	Yes
Burmese	my	Yes	Yes
Tamil	ta	Yes	Yes
Kannada	kn	Yes	Yes
Pashto	ps	Yes	Yes
Yoruba	yo	Yes	Yes
Malay	ms	Yes	Yes
Haitian Creole	ht	Yes	Yes
Nepali	ne	Yes	Yes
Sinhala	si	Yes	Yes
Catalan	ca	Yes	Yes
Malagasy	mg	Yes	Yes
Latvian	lv	Yes	Yes
Lithuanian	lt	Yes	Yes
Estonian	et	Yes	Yes
Somali	so	Yes	Yes
Tigrinya	ti	No	No
Breton	br	No	No
Fijian	fj	Yes	No
Maltese	mt	Yes	Yes
Corsican	co	Yes	Yes
Luxembourgish	lb	Yes	Yes
Occitan	oc	Yes	Yes
Welsh	cy	Yes	Yes
Albanian	sq	Yes	Yes
Macedonian	mk	Yes	Yes
Icelandic	is	Yes	Yes
Slovenian	sl	Yes	Yes
Galician	gl	Yes	Yes
Basque	eu	Yes	Yes
Azerbaijani	az	Yes	Yes
Uzbek	uz	Yes	Yes
Kazakh	kk	Yes	Yes
Mongolian	mn	Yes	Yes
Tibetan	bo	No	No
Khmer	km	Yes	No
Lao	lo	Yes	Yes
Telugu	te	Yes	Yes
Marathi	mr	Yes	Yes
Chichewa	ny	Yes	Yes
Esperanto	eo	Yes	Yes
Kurdish	ku	No	No
Tajik	tg	Yes	Yes
Xhosa	xh	Yes	No
Yiddish	yi	Yes	Yes
Zulu	zu	Yes	Yes
Sundanese	su	Yes	Yes
Tatar	tt	Yes	Yes
Quechua	qu	No	No
Uighur	ug	No	No
Wolof	wo	No	No
Tswana	tn	Yes	Yes

Additional Resources

• PyPI page
• ISO 639-1 Codes
• Github project repository
• Documentation

Authors

• Adam Pawełek
  • LinkedIn
  • Email

License

llmtranslate is licensed under the MIT License. See the LICENSE file for more details.