Build an AI engineering tool that helps developers select the most useful context for a model task when the available context window or token budget is limited.
Allow users to provide a task description, system instructions, source documents, code files, conversation history, and a maximum token budget. Split the materials into traceable sections and estimate their token usage using the selected model's tokenizer when available.
Rank context candidates by task relevance, information uniqueness, recency where relevant, dependency importance, and evidence quality. Identify duplicated explanations, obsolete details, conflicting instructions, and information that can be summarized without losing critical meaning.
Generate several context packs for comparison: maximum coverage, minimum tokens, source-first, and dependency-aware. Show what was excluded, compressed, or retained, with links back to the original material.
Include an evaluation mode that compares task performance across different context packs when a model connection is available. Never assume that shorter context automatically produces better results, and preserve critical constraints rather than removing them solely to save tokens.
0 Comments