A 311 team wants RAG answers about debris-pile complaints that arrive as handwritten photo attachments. What must their ingestion flow include?

A 311 team grounding RAG answers on debris-pile complaints that arrive as handwritten photo attachments must include OCR in the ingestion flow. Handwritten notes on photos—addresses, pile size, alley markers, pickup preferences—remain unusable pixels until OCR converts them to text the chunking and embedding stages can index. Without that step, RAG cannot cite the attachment evidence residents and inspectors actually wrote. Running translation on handwritten debris-pile complaint attachments changes language but never converts the handwritten addresses, pile-size notes, or alley markers from pixels into text, so RAG cannot retrieve and cite those specifics. Running summarization on handwritten debris-pile complaint attachments only condenses text that is already extracted but never reads the handwritten address or pile-size notes scrawled on them, so the specific details residents wrote down are still lost. Detecting layout in handwritten debris-pile complaint attachments maps regions and structure but likewise never converts the handwritten notes into text the retrieval pipeline can cite. Exam caveat: when the stem pairs RAG with handwritten municipal photo attachments, OCR ingestion is mandatory; running translation, running summarization, or detecting layout without transcription are distractors. Operational check: OCR handwritten debris attachments into the corpus with ticket IDs; spot-check address and pile descriptions against originals; run RAG evaluation questions that depend on attachment text, not just translated output, before enabling the 311 bot.