BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//pydata-amsterdam2026//speaker//UDP8AQ
BEGIN:VTIMEZONE
TZID:CET
BEGIN:STANDARD
DTSTART:20001029T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000326T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-pydata-amsterdam2026-DQXPVN@pretalx.com
DTSTART;TZID=CET:20260911T132500
DTEND;TZID=CET:20260911T135500
DESCRIPTION:RAG is the accepted paradigm to work with knowledge bases that 
 don't fit into a typical context window. But RAG is good at retrieving _on
 ly some_ relevant bits\; what do you do if there's **10M tokens of relevan
 t information** which your agent needs to act on? In this talk we explore 
 tactics to deal with 1m+ context window and evaluate whether how accuracy 
 and performance degrade at large context volumes.
DTSTAMP:20260710T154924Z
LOCATION:Main stage
SUMMARY:When RAG is not enough: Architecting for 10M+ context windows - Aza
 mat Omuraliev
URL:https://pretalx.com/pydata-amsterdam2026/talk/DQXPVN/
END:VEVENT
END:VCALENDAR