contextual-knowledge-four-sources

IN premise — summaries/2026/08/24/xu-2024-knowledge-conflicts-survey-s0-abstract.md

Created 2026-08-25T02:59:01+00:00

Contextual knowledge in LLMs comprises four sources: user prompts, multi-turn dialogue history, retrieved web documents (RAG), and tool/API outputs.

Summary

This means that everything an LLM can draw on during a live interaction comes from only four places: what you typed, the earlier messages in the thread, documents pulled in from external sources, and results returned by connected tools. That gives a finite, checkable set of inputs to audit, so if the model says something, you should be able to trace it back to one of those four sources or recognize it as coming from its pre-trained weights rather than the current context.