Context Viewer: Turning LLM Contexts into Analyzable Artifacts
Srihari Sriraman (nilenso), Michael Isaac (Carnegie Mellon University), Atharva Raykar (nilenso), Heather Miller (Carnegie Mellon University)
Engineering & Operations
Summary
A browser-based visual analytics system for exploratory analysis of LLM contexts, enabling users to inspect token usage, compare trajectories, and debug agent failures.
Description
As large language models (LLMs) move toward long-horizon tasks and multi-turn interactions, there is an increasing need to inspect and analyze the context that shapes them. However, existing systems typically present context as monolithic text transcripts or message logs in JSON. This makes it difficult to analyze its components and assess where tokens are spent, when irrelevant material accumulates, and how trajectories differ across runs. We present Context Viewer, a browser-based visual analytics system for exploratory analysis of LLM contexts. It uses an LLM-assisted pipeline to divide the context into semantically meaningful units, identify topics, and classify the units into those topics. It allows users to iteratively revise the prompts in the pipeline to refine the components and steer their exploration. And it provides interactive visualizations to filter, group, and compare contexts side-by-side. Context Viewer enables practical tasks such as analyzing agent failures, evaluating context compaction effectiveness, and studying the growth of system prompt components over time, functioning as an observability tool for context engineering.
