上下文压缩问题 #842

Activity

1. harryfan1985 commented on May 23, 2026

   

   harryfan1985

   on May 23, 2026

   Contributor

   More actions

   Synced update from PR #845: #845

   This PR includes context-compression changes relevant to this issue:

   • Replaced the prior sequential multi-chunk compression flow with a map/reduce-style flow.
   • Added bounded map-phase concurrency through map_reduce_max_parallel_requests (default: 4), avoiding unbounded parallel model requests for very large sessions.
   • Added reduce-stage token-budget protection. If all segment summaries do not fit in one reduce prompt, compression now performs bounded batched reduce passes.
   • If reduce still cannot fit safely, the compressor returns through the existing structured local fallback path instead of sending an oversized request to the model.

   Follow-up hardening was added in commit 907e2d23:

   • fix(core): harden prompt caching compression

   Verification passed:

   • cargo check -p bitfun-core
   • cargo test -p bitfun-core compression -- --nocapture
   • cargo test -p bitfun-core agentic::agents::prompt_builder -- --nocapture
   • cargo check --workspace
   • cargo test --workspace
2. GCWing commented on May 28, 2026

   

   GCWing

   on May 28, 2026

   Collaborator

   More actions

   已有部分优化，可以等新版本尝试（0.2.9，预计5.29发布），Issue暂时关闭，后续版本有问题可以继续打开讨论
3. GCWing

   closed this as completedon May 28, 2026