Reported evidence
read_research_rows chunked reader → uses → Dual row_no and chunk_no cursor
The reader advances by both row number and chunk number, allowing it to resume within a record while distinguishing chunks belonging to different records.
Read the evidence
The read_research_rows implementation first serializes each row with row_to_json(source)::text, then obtains row_no through stable ordering; each row is split into 8,192-character chunks, with at most eight chunks fetched each time. The outer cursor stores both row_no and chunk_no and continues reading after the last returned chunk. Recording only the row number would omit subsequent chunks of the same row, while recording only the chunk number would not distinguish multiple rows.Read the full research