RAG Misses French or Arabic Documents

A model can answer fluently in French or Arabic while the retrieval system still misses relevant documents in those languages. Check text extraction, embedding compatibility and cross-language retrieval separately from the language of the final answer. Research on Arabic-English retrieval bias documents this as an evaluation problem; it does not predict your application's performance.

Test the document text first

Open the extracted passage. For Arabic, inspect character order, disconnected letters and mixed left-to-right identifiers. For scanned documents, confirm that the OCR engine handled the intended language.

Search for a known phrase in the extracted text. If the phrase is absent or damaged, changing the answer language cannot repair the missing evidence.

Use the same source fact in two languages where you can verify the translation. Do not use an unreviewed machine translation as unquestioned ground truth.

Check the embedding model's instructions

Some embedding models require different prefixes for queries and passages. The multilingual E5 model card documents its prefixes, token limit and language limitations. Those requirements are model-specific, not universal settings for every embedding model.

Record the model and revision used for both indexing and querying. Verify preprocessing on both paths. An English-only query rewrite can also discard names or technical terms that matter in the original language.

Build a language matrix

Query language	Document language	Expected evidence
English	English	Known passage
French	French	Equivalent passage
Arabic	Arabic	Equivalent passage
English	French or Arabic	Cross-language match
French or Arabic	English	Cross-language match

Keep names, dates and identifiers in the test cases. A system that succeeds only on generic sentences may still fail on the documents your readers use.

Inspect retrieved passages before reviewing the generated answer. A fluent answer based on the wrong source should fail the test.

Repair the failing stage

For broken extraction, fix the parser or OCR. For missing cross-language matches, compare a suitable multilingual embedding configuration and a carefully evaluated query-translation path.

Keep original terms alongside translated variants when they identify products, people or records. Translation can alter a code or proper noun and make exact matching worse.

If retrieval works but the answer uses the wrong language, adjust the answer instruction without rebuilding the index unnecessarily. These are distinct failures.

Report performance honestly

State which language pairs and document types were tested. “Multilingual” does not mean equally accurate across every language, dialect or script.

Use the RAG overview to locate the stages, and prompt regression testing to preserve the paired cases. For Lucivo readers working across English, French and Arabic, that explicit matrix is more useful than a generic multilingual badge.

Related troubleshooting

• RAG Cannot Read a Scanned PDF: What to Check
• Change Embedding Models Without Mixing Vectors

This guide draws on the linked documentation. Examples are illustrative unless explicitly identified as measured results.