Ingestion wants to delete the layout parser and embed p...

While vision-language models offer a modern, simplified ingestion path, they sacrifice the precision of sparse retrieval for exact matches (e.g., part numbers, error codes) and span-level citations. Given the high volume and strict latency requirements, a hybrid system is best: use block extraction for the lexical and citation leg, and add page-image embeddings as a second leg, fusing them by rank. The decision should ultimately depend on the share of query traffic requiring exact alphanumeric matches.