Bug: Non-English queries never translated before embedding — broken multilingual retrieval

[anushagarg6]

Description

The new i18n system allows users to ask questions in French, Spanish, Hindi, Portuguese, and other languages. The answer_question pipeline accepts a language parameter and correctly translates the answer back to the user's language. However, the user's original question is embedded and searched against ChromaDB as-is, without first being translated to English.

Since the entire IPCC corpus and research papers are indexed in English, embedding a French or Hindi question produces a query vector that poorly matches the English document space, resulting in low-quality or empty retrieval.

translate_to_english() already existed in backend/llm/translation.py but was never called in the retrieval pipeline.

Reproduction

1. Set language to French (fr) in the UI dropdown
2. Ask: "Qu'est-ce qu'une vague de chaleur marine ?"
3. Observe: retrieval returns 0 or irrelevant chunks despite the question being valid

Impact

All non-English queries return degraded or empty answers, making the multilingual feature effectively non-functional.

Location

backend/retrieval/pipeline.py — answer_question() function.

Fix Applied (commit 2936cb5)

Added retrieval_question = translate_to_english(question) if language != "en" else question before calling query_chunks(), so the translated English version is used for embedding while the original question is preserved for history and display.