feat(runtime): 文档摄取①(抽取器 + sidecar + 装配期 resolveDocuments)(design/24 S3-A)
CI / Backend (build / vet / test / cross-compile) (push) Has been cancelled
CI / Frontend (tsc + vite build) (push) Has been cancelled

- 抽取器零新依赖纯函数:UTF-8 清洗、CSV/TSV 行数列名摘要、64K rune 安全截断;PDF/Office/二进制走哨兵错误诚实拒绝(不引解析库)。

- sidecar:<DataDir>/extracted/<存储名>.txt,首行紧凑 JSON 元数据 + 正文;临时文件 + rename 原子落定;只抽一次(删原文件仍出正文);版本不匹配/坏缓存重抽。

- 装配期 resolveDocuments(resolveImages 同位,每步对模型可见视图注入):清单恒在(解析器缺席/报错也在)、截断带续读游标、失败诚实标记 + status=unreadable、回合继续;文档部件永不抵达编码器(负例钉住)。

- 修 S2 移交的两处静默吞掉:checkpoint 比较键纳入文档引用、compactor token 估算计入文档部件(不是免费的)。

- 主会话独立验收:隔离 worktree(HEAD + 本切片)build/vet 绿、全树 59 包 0 FAIL、linux/amd64 + darwin/arm64 交叉编译绿、documents+loop 两包 167 条 PASS;变异独立复现(resolveDocuments 短路 ⇒ 3 条红,含「文档部件仍抵达下游」)。
This commit is contained in:
AskaEth committed 2026-10-04 17:22:05 +08:00
1 parent 834ef3683e
commit a8851dc0a2
12 files changed
+2008 -7

No files matched your search

+14
View File
@@ -14,6 +14,7 @@ import (
"git.yeij.top/AskaEth/Cyrene_v2/internal/persona/loader"
personareg "git.yeij.top/AskaEth/Cyrene_v2/internal/persona/registry"
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/documents"
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/embedding"
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/loop"
@@ -222,6 +223,18 @@ func (a *app) wireRuntime() error {
a.dispatcher.Invoke,
)
// 文档摄取(design/24 §2.3-①、§4):装配期把文档部件确定性替换为抽取所得
// 的正文或「仅清单」条目,文档部件因此永不抵达编码器。路径口径与 S1 的
// `POST /api/v1/documents` 同源 —— 原文在 `<DataDir>/uploads/documents`
// (gateway 的 documentSubDir),抽取结果缓存到 `<DataDir>/extracted`。
// 这里复用同一个 resolveDataDir 结果,两处指向同一棵目录树。
docDataDir := resolveDataDir(a.cfg)
docResolver := documents.NewResolver(documents.Config{
DataDir: docDataDir,
UploadsDir: filepath.Join(docDataDir, "uploads", "documents"),
Log: a.log,
})
capDecider := a.buildCapacity()
// 人格情绪观测器(design/01 §2)在此构建,因为它是 loop 的 turn 边界依赖:
// 每个 turn 落定后,该 turn 绑定人格的情绪机与好感度都会推进,返回的心情标签
@@ -271,6 +284,7 @@ func (a *app) wireRuntime() error {
WorkingMemory: memory.NewWorkingMemory(0),
OnStream: a.streamChunk,
Vision: visionResolver,
DocumentResolver: docResolver,
Thinking: thinking,
ReasoningEffort: reasoningEffort,
ThinkingShowToUser: thinkingShow,