Unverified50% confidenceFactExact time
SR-SWA limits the number of active reference tokens to a fixed capacity, so that the per-step attention computation cost does not grow with document length.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRate Limiting refers to a system restricting the number of requests a user can make within a unit of time.65% similarUnverifiedCached input tokens typically enjoy a discount of 25%–50% off the original price, as there is no need to recompute embeddings and attention weights—only the precomputed KV vectors need to be loaded from cache.65% similarUnverifiedIn frameworks where prompt caching is already enabled, the unit price for cache hits is only 1/50 that of new input, so cost savings are limited when the volume of judgments is not large.65% similarUnverifiedWhen the number of project files exceeds the token limit that the model can process in a single pass, manually scoping relevant files using the @ marker can reduce irrelevant responses and lower token consumption.64% similarUnverifiedProduction-grade Agents require setting token budget caps and implementing rate limiting to avoid unexpectedly high API costs caused by recursive calls, invalid retries, or excessively large context windows.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1113972API
curl https://kongchang.com/api/v1/knowledge/claims/1113972MCP
get_claim(id=1113972)