panda/agent_jrxml
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: OCR fidelity scoring reform — prevent false fail from language-mismatched field names

Browse Source 
Root cause (from review): field_coverage compared English JRXML field names
against Chinese OCR field names with set intersection — always zero. Combined
with 0.5 weight in score formula, caused valid JRXML (XSD pass, 82% element
coverage) to score 0.41 < 0.5 → fail → correction loop → progressive destruction.

Changes:
- Scoring weight: element_coverage 0.8 + field_coverage 0.2 (was 0.5+0.5)
- Validate node: only fail on fidelity when BOTH score<0.5 AND element_coverage<0.4
- Field name regex: \w+ → [^"]+ to support non-ASCII field names
- Field matching: also try _sanitize_field_name conversion (Chinese→_uXXXX_)
- correction.md: namespace check always active, not conditional on error keywords
This commit is contained in:
panda
2026-05-24 15:36:40 +08:00
parent 4e14334030
commit 2d5183d2bd
2 changed files with 17 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
agent/nodes.py
+16 -9
View File
@@ -1207,8 +1207,8 @@ def _check_ocr_fidelity(jrxml: str, state: dict) -> dict:
else:
element_coverage = 1.0
# 2. 字段名覆盖
jrxml_fields = set(re.findall(r'<field name="(\w+)"', jrxml))
# 2. 字段名覆盖（英文字段名 vs OCR 中文字段名天然不匹配，权重降低）
jrxml_fields = set(re.findall(r'<field name="([^"]+)"', jrxml))
ocr_field_names = set()
ocr_fields = ocr_result.get("fields", []) if isinstance(ocr_result, dict) else []
for f in ocr_fields:
@@ -1219,10 +1219,15 @@ def _check_ocr_fidelity(jrxml: str, state: dict) -> dict:
if ocr_field_names and jrxml_fields:
matched = jrxml_fields & ocr_field_names
field_coverage = len(matched) / max(len(ocr_field_names), 1)
unmatched = ocr_field_names - jrxml_fields
if unmatched:
sample = list(unmatched)[:8]
# 尝试通过 _sanitize_field_name 转义后匹配（中文→_uXXXX_）
sanitized_ocr = {_sanitize_field_name(n) for n in ocr_field_names}
matched_via_sanitize = jrxml_fields & sanitized_ocr
all_matched = matched | matched_via_sanitize
field_coverage = len(all_matched) / max(len(ocr_field_names), 1)
still_unmatched = {n for n in ocr_field_names
if n not in jrxml_fields and _sanitize_field_name(n) not in jrxml_fields}
if still_unmatched:
sample = list(still_unmatched)[:8]
issues.append(f"OCR 字段未在 JRXML 中声明: {', '.join(sample)}")
elif ocr_field_names and not jrxml_fields:
field_coverage = 0.0
@@ -1247,8 +1252,9 @@ def _check_ocr_fidelity(jrxml: str, state: dict) -> dict:
f"OCR 布局分析有 {ocr_columns} 列"
)
# 综合评分
score = round(field_coverage * 0.5 + element_coverage * 0.5, 3)
# 综合评分（元素覆盖为主权重 0.8，字段覆盖为辅助 0.2）
# 英文字段名 vs 中文 OCR 字段名天然不匹配是预期行为，字段覆盖仅作参考
score = round(element_coverage * 0.8 + field_coverage * 0.2, 3)
return {
"score": score,
"field_coverage": round(field_coverage, 3),
@@ -1291,7 +1297,8 @@ def validate(state: AgentState) -> Dict:
if fidelity["issues"]:
if state["status"] == "pass":
# XSD 通过但内容保真度不足 → 降级为 fail
if fidelity["score"] < 0.5:
# 需要 score < 0.5 且 element_coverage < 0.4（字段名语言不匹配不应单独导致 fail）
if fidelity["score"] < 0.5 and fidelity.get("element_coverage", 0) < 0.4:
state["status"] = "fail"
state["error_msg"] = (
f"[内容保真度不足] 得分 {fidelity['score']:.2f}/1.0。"
prompts/correction.md
+1 -1
View File
@@ -8,7 +8,7 @@
- 如果当前 JRXML 内容为空或过短（<200 字符），请根据下方提供的 OCR 识别数据和布局 schema 重新生成完整的 JRXML，而非输出一个占位桩。
- 如果错误是"字段 'field_N' 未在 <field> 部分声明"，**必须**为每个缺失的 field_N 添加 `<field name="field_N" class="java.lang.String"/>` 声明。这些是占位字段，不可删除。同时确保所有 $F{{field_N}} 引用都有对应的 <field> 声明。
- 如果错误是"字段 'field_N' 未在 <field> 部分声明"且有 OCR 字段数据，尝试将 $F{{field_N}} 替换为 OCR 中对应的真实字段名（如 $F{{invoice_code}}），同时更新 <field> 声明和所有引用。
- **如果错误包含 "No matching global declaration available for the validation root" 或 "namespace" 或 "命名空间"**，说明命名空间 URL 错误。正确的根元素格式必须为：`<jasperReport xmlns="http://jasperreports.sourceforge.net/jasperreports" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://jasperreports.sourceforge.net/jasperreports http://jasperreports.sourceforge.net/xsd/jasperreport.xsd">`。删除所有 ns0: 前缀，删除所有 `xmlns:ns0` 声明，删除所有元素标签上的 `ns0:` 前缀。
- **始终检查并修复命名空间**：正确的根元素格式必须为：`<jasperReport xmlns="http://jasperreports.sourceforge.net/jasperreports" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://jasperreports.sourceforge.net/jasperreports http://jasperreports.sourceforge.net/xsd/jasperreport.xsd">`。删除所有 ns0: 前缀，删除所有 `xmlns:ns0` 声明，删除所有元素标签上的 `ns0:` 前缀。
当前 JRXML（带错误）：
{current_jrxml}