Pulse grounded search assigns story headlines to wrong URL chunks. The Vertex AI grounded search returns N chunks with resolved URLs. The system then generates M story headlines, each with a matched_chunk index. But the matched_chunk indices are assigned sequentially (0, 1, 2) rather than semantically matched to the correct chunk. Result: headline says 'Fitch Ratings Warns of Shrinking Real Incomes' but links to a KPMG article about CBP de minimis customs rules. The correct chunk (economynext.com about Fitch) was at index 3, and the correct cookout chunk (fb.org Farm Bureau) was at index 4, but stories 1 and 2 got indices 1 and 2 instead. Visible in manifest.json macro_trace.data: chunks_resolved has 5 entries, stories has 3 entries with matched_chunk values 0,1,2 — should have been 0,3,4.
The bug is in the story-to-chunk matching logic in the grounded search pipeline. When the LLM generates story headlines from the grounded search chunks, each story needs to be matched to the chunk whose content actually corresponds to that headline. Currently the matched_chunk index appears to be assigned sequentially (story 0 -> chunk 0, story 1 -> chunk 1, etc.) rather than by semantic match. Fix: after generating the story headlines, use the chunk titles/URLs/content to match each story to the correct chunk — either by having the LLM output the correct chunk index, or by doing a post-hoc semantic similarity match between story headline and chunk title/content. Alternatively, ensure the LLM prompt that generates stories includes the chunk index as a required output field and validates it against the available chunks.