← All questions
Retrieval & EmbeddingsHard· Indexing· Table Parsing· Retrieval Strategy

Design the retrieval unit for a 200-page filing contain...

A single approach is insufficient for complex documents. Prose should be processed using standard chunking rules. Tables should be indexed as atomic units, but you must implement a row-atomic fallback that repeats the table header for each row to maintain context. Finally, create a surrogate summary for each table so that they can be retrieved via semantic search using their names or descriptions. Note that implementing the row-atomic fallback will result in approximately a 39% increase in token usage.

Code playground (Python)