RAG Document API

Task

A document question-answering API retrieves passages from uploaded PDFs and sends them to a model as evidence. Build the context builder that turns retrieved passages into the model’s bounded evidence context, with the page-level sources the answer API cites. Ingestion, retrieval, the LangGraph workflow and the model adapters are supplied and run locally.

Requirements

1. Return exactly {context, sources}. Format each included passage as "[id] document=document_id page=page", a newline, then its stripped text. Join sections with two newlines.
2. max_chars is an integer from 1 to 12000, not a boolean (ValueError containing max_chars). Headers and separators count against it. Skip a whole passage that does not fit and continue with later, smaller ones; never truncate a passage.
3. passages is a list of objects with a nonblank string id and document_id, a positive integer page (not a boolean), and string text. Reject anything else with ValueError containing passage.
4. Preserve retrieval order, keep only the first occurrence of each id, and skip blank text. sources lists {id, document_id, page} for each included section. Do not mutate the input.