BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//sf26-agenda//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
UID:SESS-111@sf26-agenda
DTSTAMP:20261020T151709Z
DTSTART:20261105T173000Z
DTEND:20261105T181500Z
SUMMARY:When the RAG-Time Never Stops
LOCATION:Room 2000
DESCRIPTION:Benjamin Gamble\n\nMost RAG systems still behave as if the worl
 d moves in polite little batches. Crawl the corpus\, chunk the documents\,
  vectorise the lot\, load the store\, and hope nobody asks a question befo
 re the next refresh. That may work for static PDFs. It does not work when 
 the facts are changing under your feet. This talk is about treating retrie
 val as a live data problem. Or\, to put it another way: if the RAG-time ne
 ver stops\, your knowledge system needs to keep playing. Modern AI systems
  need continuous ingest\, continuous vectorisation\, and continuous recall
  over both hot and cold data. New documents\, operational events\, user pr
 ompts\, market signals\, policy updates\, and application changes all need
  to enter the system at the right speed. Some data belongs on the fast pat
 h. Some belongs in the archive. Some needs re-ranking\, enrichment\, graph
  context\, or direct lookup before it is useful. The trick is not to make 
 everything instant. The trick is knowing what must be fresh\, what must be
  trusted\, and where it came from. We will look at how to build multi-spee
 d RAG pipelines using streaming platforms\, stream processing\, vector sto
 res\, search systems\, graph databases\, and event-driven enrichment. The 
 focus is not one magic store or one “easy winner” architecture. It is 
 how to keep recall adaptable while preserving data lineage\, freshness\, a
 nd operational control. Because stale recall is not just a quality issue. 
 In financial services\, logistics\, risk\, news\, and operational AI\, sta
 le recall can mean wrong decisions made confidently. And once the system s
 tarts making decisions from yesterday’s tune sheet\, even The Entertaine
 r cannot save you. This is RAG as feature engineering: live\, traceable\, 
 adaptive\, and always in time. Listeners will learn: how to design RAG sys
 tems as continuous data products rather than batch indexing jobs. how to s
 eparate hot\, warm\, and cold retrieval paths by freshness\, cost\, and tr
 ust. why lineage matters for AI recall: what was known\, when it was known
 \, and which answer depended on it. practical patterns for building adapti
 ve retrieval pipelines that keep pace with changing data.
URL:https://sf26.rmoff.net/s/when-the-rag-time-never-stops-111
END:VEVENT
END:VCALENDAR