Machine Translation with LLMs

LLM Translation Capabilities Modern LLMs are highly capable translators, often matching or exceeding specialized translation models.

Basic Translation

def translate(text: str, target_lang: str) -> str:
    return llm.generate(f"""
Translate the following text to {target_lang}.
Only output the translation, nothing else.

Text: {text}
Translation:
    """)

Advanced Translation

With Context

def translate_with_context(text: str, target_lang: str, context: str) -> str:
    return llm.generate(f"""
Translate to {target_lang}, considering this context:
Context: {context}

Text to translate: {text}
Translation:
    """)

Style Preservation

def translate_preserve_style(text: str, target_lang: str) -> str:
    return llm.generate(f"""
Translate to {target_lang} while preserving:
- Tone (formal/informal)
- Technical terminology
- Cultural nuances

Original: {text}
Translation:
    """)

Multilingual RAG

Cross-Lingual Retrieval

def multilingual_rag(query: str, query_lang: str, doc_lang: str) -> str:
    # Translate query to document language
    translated_query = translate(query, doc_lang)

    # Retrieve in document language
    docs = retrieve(translated_query)

    # Generate response in query language
    response = llm.generate(f"""
Answer in {query_lang} using these documents:
{docs}

Question: {query}
    """)
    return response

Multilingual Embeddings Use cross-lingual embedding models:

Specialized Translations

Technical Documents

glossary = {"API": "API", "endpoint": "punto final"}

def translate_technical(text: str, glossary: dict) -> str:
    return llm.generate(f"""
Translate to Spanish using this glossary:
{glossary}

Keep terms in glossary consistent.
Text: {text}
    """)

Localization Beyond translation: adapt for local conventions:

• Date formats
• Currency
• Cultural references
• Measurement units

Quality Considerations