重排报错

[jmzd315221]

在小白x 的监控里看到的，Rerank API 的输入 Token 数（8194）超出了模型允许的最大上下文长度（8192）,相同的错误日志连报了三次

原日志：

Error: Rerank API 400: {"code":20015,"message":"Query is too long. Please provide a shorter query.","data":null}
    at rerank (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/vector/llm/reranker.js:125:19)
    at async runBatch (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/vector/llm/reranker.js:216:37)
    at async worker (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/vector/llm/reranker.js:243:13)
    at async Promise.all (index 0)
    at async rerankChunks (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/vector/llm/reranker.js:246:5)
    at async locateAndPullEvidence (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/vector/retrieval/recall.js:821:22)
    at async recallMemory (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/vector/retrieval/recall.js:1492:61)
    at async buildVectorPromptText (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/generate/prompt.js:1390:24)
    at async EventEmitter.handleGenerationStarted (http://127.0.0.1:8765/scripts/extensions/third-party/LittleWhiteBox/modules/story-summary/story-summary.js:2958:19)
    at async EventEmitter.emit (http://127.0.0.1:8765/lib/eventemitter.js:306:17)

截图：

[Hhhhenrry-HA]

目前取最近3楼消息拼 query 做 rerank，容易超 token 的原因更可能是 query 拼接时混进了大块不需要参与 rerank 的内容。常见两类：

1. user 那边用了剧情规划等长文本XML 包裹块
2. AI 回复里带了状态栏、XML 包裹块、工具输出等
   这些如果直接混进 rerank query，8192 很容易满。设置里是有清洗项的（比如 、状态栏等），可以先看下是不是这块没配你的卡的特殊标签？导致这些内容没被洗掉。