Prefetching
In the context of LLM agents, a data-access strategy in which an external orchestration layer calls data-source APIs before the model begins inference and injects the results into the prompt as context. It emphasizes input consistency and controllability.
Timeline (last 90 days)
For high-frequency, small-field scenarios, prefetching is often more economical; for low-frequency, large-field or long-tail queries, MCP's on-demand nature can better compress wasted overhead.
A hybrid approach is suggested: prefetching a small set of core fields plus on-demand MCP supplementation, to balance controllability and flexibility.
A practical criterion is: if a field is used in more than 80% of reasoning, prefetch it; if its usage rate is low and its size is large, leave it to MCP.
In prefetch mode, if 20 fields are prefetched but only 5 are actually used, the redundancy ratio is as high as 75%.
In the LLM Agent context, prefetch mode refers to an external orchestration layer calling data-source APIs in advance before the model begins reasoning, then formatting the results and injecting them into the prompt as context.
Even an Oracle prefetcher with foreknowledge of future requests never outperforms the no-prefetch approach on the migration traffic metric.
Common prefetching strategies include sequential-detection-based streaming prefetch, B-Tree prefetch based on index leaf-node traversal paths, and machine learning prefetch models trained on historical query logs.
The key challenge in prefetching is the trade-off between accuracy and bandwidth overhead: aggressive prefetching consumes bandwidth and cache capacity, while conservative prefetching fails to hide remote I/O latency.
All Facts (8)
For high-frequency, small-field scenarios, prefetching is often more economical; for low-frequency, large-field or long-tail queries, MCP's on-demand nature can better compress wasted overhead.
50%UnverifiedA hybrid approach is suggested: prefetching a small set of core fields plus on-demand MCP supplementation, to balance controllability and flexibility.
50%UnverifiedA practical criterion is: if a field is used in more than 80% of reasoning, prefetch it; if its usage rate is low and its size is large, leave it to MCP.
50%UnverifiedIn prefetch mode, if 20 fields are prefetched but only 5 are actually used, the redundancy ratio is as high as 75%.
50%UnverifiedIn the LLM Agent context, prefetch mode refers to an external orchestration layer calling data-source APIs in advance before the model begins reasoning, then formatting the results and injecting them into the prompt as context.
50%UnverifiedEven an Oracle prefetcher with foreknowledge of future requests never outperforms the no-prefetch approach on the migration traffic metric.
50%UnverifiedCommon prefetching strategies include sequential-detection-based streaming prefetch, B-Tree prefetch based on index leaf-node traversal paths, and machine learning prefetch models trained on historical query logs.
50%UnverifiedThe key challenge in prefetching is the trade-off between accuracy and bandwidth overhead: aggressive prefetching consumes bandwidth and cache capacity, while conservative prefetching fails to hide remote I/O latency.
50%