BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//pydata-amsterdam2026//speaker//JHJX93
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-amsterdam2026-7SGSFY@pretalx.com
DTSTART;TZID=CET:20260910T103500
DTEND;TZID=CET:20260910T110500
DESCRIPTION:This talk is about the art of forgetting\, a design pattern for
  long-context LLM pipelines where you drop the noise from the context on p
 urpose\, and the model calls a retrieval tool to pull back anything it act
 ually needs. You pay for half the tokens\, and the model capacity to answe
 r is maintain\, sometimes even improved.\n\nI'll walk through the pattern 
 end-to-end in about forty lines of Python\, content-type routing (image\, 
 code\, plain text\, etc)\, query-aware relevance scoring\, compressed-stor
 e markers\, and retrieve-tool wiring. \n\nThen\, the part that matters mos
 t to anyone actually shipping this: how to test whether forgetting-on-purp
 ose helps your workload. \n\nI'll share a reproducible evaluation recipe\,
  including the naive-baseline control that separates what works and what d
 oesn't\, and the checks that catch benchmarks measuring the wrong compress
 or.\n\nYou'll leave with a provider-neutral architectural pattern you can 
 drop into any Python LLM pipeline\, and a reusable methodology for decidin
 g whether it's worth it on your own production traces. \n\nReference imple
 mentation: the open-source Headroom library and my benchmark repository wi
 th all code and evaluation assets will be public.
DTSTAMP:20260710T150510Z
LOCATION:Room 2 (350)
SUMMARY:The Art of Forgetting: A Compress-Retrieve Pattern for Long-Context
  LLMs - Arian Pasquali
URL:https://pretalx.com/pydata-amsterdam2026/talk/7SGSFY/
END:VEVENT
END:VCALENDAR