iomgaa
bb621cad4d
Zip construction and member reads now share one narrow boundary so corrupt BZIP2, LZMA and version fields cannot leak native exceptions. Non-success statuses are read before classification, text-bearing blocks require lines, the size cap and reasoning contract are pinned by tests that fail under mutation, and the per-page timing claim is corrected to the observed range.
18 lines
1.8 KiB
Markdown
18 lines
1.8 KiB
Markdown
---
|
||
type: metric
|
||
node_id: metric:ocr-call-success
|
||
title: "OCR 治理调用成功率与错误分类分布"
|
||
date: 2026-07-21
|
||
---
|
||
|
||
# OCR 治理调用成功率与错误分类分布
|
||
|
||
- **数据源**: `llm_calls` 表(复用,无新表;M3 设计 §7)。OCR 行特征: `provider="monkey"`、tokens=0、cost=NULL、ttft/inter=NULL、cache_hit=0;`messages` 为 `<ocr:text|layout|document image_bytes=N>` 占位(图像 bytes 绝不入库),`response` 为 text 截断 200 字 / `<elements n=X pages=Y>` 摘要。
|
||
- **埋点位置**(唯一路径,禁止旁路): `src/polygateway/ocr.py` OcrClient 治理循环内经 `TelemetryEmitter.emit_attempt`(单 helper 铁律),每次尝试(成功/失败/取消)必录;写失败 warning 降级不冒泡。
|
||
- **指标口径**:
|
||
- 调用成功率 = `error IS NULL` 行数 / 总行数(按 provider="monkey" 过滤,窗口按 run/日);
|
||
- 错误分类分布 = `error` 非空行按异常类名前缀(Transient/SourceDead/RequestRejected/ResultInvalid/CircuitOpen/AllSourcesExhausted)归组计数;
|
||
- 端点维度经 `messages` 占位前缀(`ocr:text` / `ocr:layout` / `ocr:document`)区分,或直接用 `operation` 列(`recognize_text` / `parse_layout` / `parse_document`);文档行的 `response` 摘要为 `<document elements=N pages=M md_chars=K>`,**不含正文**;
|
||
- 延迟 = `latency_ms` 分位数(text、layout、document 分列——后两者含两段 HTTP,显著更慢;document 随页数增长,约 1.5–4 s/页[2026-10-21 实测:2 页 3.5–8 秒、4 页 6.2–6.4 秒;样本极少、无争用,非 SLA],不要与 layout 混算分位数)。
|
||
- **评估基线**: 待 P7 OCR soak 首跑后建立(健康双真源池预期成功率 ≥98%,故障池口径阈值在 P7 计划冻结);单调用延迟基线以 integration 实测为初值(2026-07-21 手测: /ocr/text 约 3-5s,/parse 约 5-15s)。
|