Part B: How the current Climate Chatbot handles languages

The Climate Academy knowledge base is primarily stored as English chunks. The current multilingual approach can therefore be understood as a translation-first RAG pipeline:

1. A user asks a question in one of the supported languages.
2. The system detects the language.
3. The question is translated into English.
4. An embedding is generated from the English question.
5. ChromaDB searches the English Climate Academy chunks.
6. The most relevant English passages are retrieved.
7. The LLM uses those passages to create a grounded English response.
8. The answer is translated back into the user’s original language.
9. The user receives the final answer with source information.

Advantages of translation-first retrieval

• Single knowledge base: The project does not need a separate full copy of the book for every language.
• Easier maintenance: A corrected English source can serve all supported languages.
• Simpler architecture: The same chunking, embedding, ChromaDB, and citation workflow can be reused.
• Lower storage requirements: The vector database stores one main set of chunks rather than eight translated collections.

Challenges of multilingual AI

Translation errors

One mistranslated phrase can change the search. A word meaning ‘mitigation’ might be rendered as a general word for ‘reduction,’ causing the retriever to miss the intended climate concept.

Cultural differences

Examples that make sense in one country may feel strange or inappropriate in another. A useful explanation of household energy use should consider local housing, transport, food, and climate.

Localization is more than translation. It adapts communication while preserving scientific accuracy.

Scientific terminology

Technical words may have several translations, no widely accepted translation, or a familiar local abbreviation. The system must handle terms such as carbon sequestration, radiative forcing, tipping point, and climate justice consistently.

Low-resource languages

Some languages have far less digital text available for training. Models may therefore understand them less reliably, produce less natural answers, or fail to recognize regional varieties.

Evaluation challenges

A fluent answer can still retrieve the wrong passage or alter the scientific meaning. Evaluation requires people who understand:

• The language
• The climate topic
• The local context
• The quality of the cited evidence

Automatic translation scores alone are not enough.

Language-specific bias

A model may be more cautious, detailed, polite, or accurate in one language than another. It may reproduce stereotypes associated with particular regions or treat English sources as automatically more authoritative.

Fairness testing should compare equivalent questions across all eight supported languages.