Chunking structured documents

A retrieval system can only return what it stored, and it stores chunks. When a question about subletting retrieves "Provided that this clause shall not apply to a sharing of occupation with a group company" without the covenant it qualifies, or retrieves the covenant without the proviso, the agent answers from half a clause. In legal documents, where a later sub-clause can reverse an earlier one, how you cut is as consequential as how you embed.

This lesson starts where Document Ingestion Pipelines stops. Ingest hands over each document as typed blocks with a stable block_id, a page, and a section_path such as ["Schedule 4", "Part 2", "Service charge"], plus document-level metadata (matter, tenant, document type, status). Parsing, OCR and extraction are its job. Deciding what unit of text gets embedded and returned is yours.