How do you chunk non-English legal docs for RAG agents?

[Isaac24Karat]

Standard chunking often breaks sentence boundaries or inserts mid-paragraph splits in Hebrew and German legal docs.

Anyone tried:

• Per-language tokenizers?
• Sentence-split then re-merge by token count?
• Manual section mapping?

Would love to hear real-world methods that worked.

[Isaac24Karat]

Follow-up after testing different chunking strategies on Hebrew legal docs:

• Standard RecursiveCharacterTextSplitter still breaks context mid-sentence too often.
• Better results came from using sentence segmentation with spaCy (Hebrew-trained) → then merging small chunks under 250 tokens.
• Added metadata anchors (e.g., סעיף, פרק) as natural chunk boundaries.

Next: I’ll explore overlap tuning vs embedding window width to reduce hallucination.
Happy to share test script if useful.