context-optimization

Context optimization extends effective context length without increasing model size by compressing tokens, utilizing KV-cache, masking irrelevant observations, compacting summaries, and partitioning work across sub-agents to reduce costs and preserve critical information.

You can implement KV-cache optimization and observation masking by activating the relevant optimization techniques within the skill, which requires Python scripting and model-specific knowledge to reuse Key/Value tensors and hide irrelevant outputs after their purpose is served.

Yes, context optimization requires Python scripting and model-specific knowledge to properly implement advanced techniques like compaction, context partitioning, and KV-cache optimization for large documents and long conversations.

The best way to manage large documents and long conversations is applying context compression, compaction, and partitioning techniques to optimize token usage and reduce costs while maintaining the effective context length and performance of the model.

You should use context partitioning to split work across sub-agents for large contexts, whereas compaction is used to summarize accumulated context to preserve critical information, reducing token usage in long-running systems.