opensource/mateclaw
2
0
Fork 0
mirror of https://gitee.com/mateos/mateclaw.git synced 2026-09-13 03:13:41 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix(pdf,chat): fix OCR trigger misjudgment, false success, and duplicate image injection

Browse Source
This commit is contained in:
matevip 2026-04-06 08:18:04 +08:00
parent d8cf5acf0f
commit 7bcdbf4523
2 changed files with 89 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

117
mateclaw-server/src/main/java/vip/mate/tool/builtin/DocumentExtractTool.java
View File

@ -153,6 +153,10 @@ public class DocumentExtractTool {
// ==================== PDF 提取链 ====================
private ExtractedContent extractPdf(Path path, String options, List<String> attempts) throws Exception {
// 先获取 PDF 真实页数（不依赖提取出的文本长度反推）
int realPageCount = getPdfPageCount(path);
log.info("[DocumentExtract] PDF 真实页数: {} ({})", realPageCount, path.getFileName());
String bestContent = null;
String bestMethod = null;
@ -160,12 +164,13 @@ public class DocumentExtractTool {
long t0 = System.currentTimeMillis();
String content = tryPdftotext(path, options);
if (content != null && !content.isBlank()) {
int pages = estimatePages(content);
if (!needsOcr(content, pages)) {
if (!needsOcr(content, realPageCount)) {
attempts.add("pdftotext: 成功 (" + (System.currentTimeMillis() - t0) + "ms)");
return new ExtractedContent(content, "pdftotext", pages);
return new ExtractedContent(content, "pdftotext", realPageCount > 0 ? realPageCount : estimatePages(content));
}
attempts.add("pdftotext: 文本过少 (每页 " + charsPerPage(content, pages) + " 字符)，可能是扫描版");
double perPage = realPageCount > 0 ? (double) content.strip().length() / realPageCount : 0;
attempts.add("pdftotext: 文本过少 (总 " + content.strip().length() + " 字符, "
+ realPageCount + " 页, 每页 " + String.format("%.0f", perPage) + " 字符)，可能是扫描版");
bestContent = content;
bestMethod = "pdftotext";
} else {
@ -176,10 +181,9 @@ public class DocumentExtractTool {
long t1 = System.currentTimeMillis();
content = tryPythonPdfExtractor(path, options);
if (content != null && !content.isBlank()) {
int pages = estimatePages(content);
if (!needsOcr(content, pages)) {
if (!needsOcr(content, realPageCount)) {
attempts.add("python_pdf: 成功 (" + (System.currentTimeMillis() - t1) + "ms)");
return new ExtractedContent(content, "python_pdfplumber", pages);
return new ExtractedContent(content, "python_pdfplumber", realPageCount > 0 ? realPageCount : estimatePages(content));
}
attempts.add("python_pdf: 文本过少");
if (bestContent == null || content.strip().length() > bestContent.strip().length()) {
@ -194,10 +198,9 @@ public class DocumentExtractTool {
long t2 = System.currentTimeMillis();
content = extractPdfWithJava(path);
if (content != null && !content.isBlank()) {
int pages = estimatePages(content);
if (!needsOcr(content, pages)) {
if (!needsOcr(content, realPageCount)) {
attempts.add("java_pdf: 成功 (" + (System.currentTimeMillis() - t2) + "ms)");
return new ExtractedContent(content, "java_pdfbox", pages);
return new ExtractedContent(content, "java_pdfbox", realPageCount > 0 ? realPageCount : estimatePages(content));
}
attempts.add("java_pdf: 文本过少");
if (bestContent == null || content.strip().length() > bestContent.strip().length()) {
@ -211,10 +214,11 @@ public class DocumentExtractTool {
// 4. OCR fallback（扫描版/照片型 PDF）
log.info("[DocumentExtract] 文本提取不足，尝试 OCR: {}", path.getFileName());
long t3 = System.currentTimeMillis();
content = tryOcrExtract(path, attempts);
if (content != null && !content.isBlank()) {
attempts.add("ocr_tesseract: 成功 (" + (System.currentTimeMillis() - t3) + "ms)");
return new ExtractedContent(content, "ocr_tesseract", estimatePages(content));
OcrResult ocrResult = tryOcrExtract(path, attempts);
if (ocrResult != null && ocrResult.text != null && !ocrResult.text.isBlank()) {
attempts.add("ocr_tesseract: 成功 (" + (System.currentTimeMillis() - t3) + "ms, "
+ ocrResult.successPages + "/" + ocrResult.totalPages + " 页识别成功)");
return new ExtractedContent(ocrResult.text, "ocr_tesseract", ocrResult.totalPages);
}
// attempts 已由 tryOcrExtract 内部记录失败原因
@ -222,34 +226,72 @@ public class DocumentExtractTool {
if (bestContent != null) {
log.warn("[DocumentExtract] OCR 不可用，返回部分文本结果: method={}, length={}",
bestMethod, bestContent.strip().length());
return new ExtractedContent(bestContent, bestMethod + "_partial", estimatePages(bestContent));
int pages = realPageCount > 0 ? realPageCount : estimatePages(bestContent);
return new ExtractedContent(bestContent, bestMethod + "_partial", pages);
}
throw new Exception("所有 PDF 提取方法都失败（包括 OCR）");
}
/**
* 判断提取到的文本是否太少、需要尝试 OCR。
* 基于字符密度（每页平均字符数）而非总字符数，避免误判短票据/证书类 PDF。
* 获取 PDF 真实页数。优先用 pdfinfo（Poppler 自带），失败则用 Python pypdf。
* 不依赖提取出的文本长度反推，避免扫描版 PDF 误判页数。
*/
private boolean needsOcr(String text, int estimatedPages) {
private int getPdfPageCount(Path path) {
// 方法 1: pdfinfo（Poppler 附带工具）
try {
String output = executeCommand(List.of("pdfinfo", path.toString()));
if (output != null) {
for (String line : output.split("\n")) {
if (line.toLowerCase().startsWith("pages:")) {
String num = line.substring(6).strip();
return Integer.parseInt(num);
}
}
}
} catch (Exception e) {
log.debug("[DocumentExtract] pdfinfo 获取页数失败: {}", e.getMessage());
}
// 方法 2: Python pypdf
try {
String script = "import sys; from pypdf import PdfReader; print(len(PdfReader(sys.argv[1]).pages))";
String result = tryPythonScript(script, path.toString());
if (result != null && !result.isBlank()) {
return Integer.parseInt(result.strip());
}
} catch (Exception e) {
log.debug("[DocumentExtract] pypdf 获取页数失败: {}", e.getMessage());
}
return 0; // 未知页数
}
/**
* 判断提取到的文本是否太少、需要尝试 OCR。
* 使用真实页数（来自 getPdfPageCount）计算字符密度，不再依赖 estimatePages 反推。
* 页数未知（0）时，只看总字符数。
*/
private boolean needsOcr(String text, int realPageCount) {
if (text == null || text.isBlank()) return true;
String stripped = text.strip();
if (stripped.length() < 20) return true;
double perPage = charsPerPage(stripped, estimatedPages);
if (realPageCount <= 0) {
// 页数未知时回退到总字符数判定（保守阈值）
return stripped.length() < 100;
}
double perPage = (double) stripped.length() / realPageCount;
// 正常文本 PDF 每页至少数百字符；每页不到 30 字符大概率是扫描版
return perPage < 30;
}
private double charsPerPage(String text, int pages) {
return (double) text.strip().length() / Math.max(1, pages);
}
/** OCR 结果（含成功/失败页数统计） */
private record OcrResult(String text, int totalPages, int successPages) {}
/**
* OCR 提取：pdftoppm 转图片 + tesseract 识别文字。
* 仅依赖 Poppler（pdftoppm）和 tesseract 系统命令，不引入新 Python 依赖。
* 只有至少一页真正识别到文本时才返回有效结果，全部失败返回 null。
*/
private String tryOcrExtract(Path pdfPath, List<String> attempts) {
private OcrResult tryOcrExtract(Path pdfPath, List<String> attempts) {
Path tempDir = null;
try {
long startTime = System.currentTimeMillis();
@ -290,9 +332,11 @@ public class DocumentExtractTool {
return null;
}
// Step 4: 对每页执行 OCR
// Step 4: 对每页执行 OCR，统计成功/失败页数
long t2 = System.currentTimeMillis();
StringBuilder ocrText = new StringBuilder();
int successPages = 0;
int failedPages = 0;
for (int i = 0; i < pageFiles.length; i++) {
ocrText.append("--- Page ").append(i + 1).append(" ---\n");
try {
@ -305,17 +349,30 @@ public class DocumentExtractTool {
cmd.add(langParam);
}
String pageText = executeCommand(cmd);
ocrText.append(pageText != null ? pageText.trim() : "").append("\n\n");
if (pageText != null && !pageText.isBlank()) {
ocrText.append(pageText.trim()).append("\n\n");
successPages++;
} else {
ocrText.append("[空白页]\n\n");
failedPages++;
}
} catch (Exception e) {
log.warn("[DocumentExtract] OCR: 页面 {} tesseract 失败: {}", i + 1, e.getMessage());
ocrText.append("[OCR 失败]\n\n");
failedPages++;
// 不写 [OCR 失败] 占位，避免被上层误判为有效文本
}
}
log.info("[DocumentExtract] OCR: tesseract 处理 {} 页耗时 {}ms，总耗时 {}ms",
pageFiles.length, System.currentTimeMillis() - t2, System.currentTimeMillis() - startTime);
log.info("[DocumentExtract] OCR: tesseract 处理 {} 页 (成功 {}, 失败 {}), 耗时 {}ms, 总耗时 {}ms",
pageFiles.length, successPages, failedPages,
System.currentTimeMillis() - t2, System.currentTimeMillis() - startTime);
String result = ocrText.toString().trim();
return result.isEmpty() ? null : result;
// 只有至少一页真正识别到文本时才算成功
if (successPages == 0) {
attempts.add("ocr: tesseract 运行但所有 " + pageFiles.length + " 页均未识别到文本");
return null;
}
return new OcrResult(ocrText.toString().trim(), pageFiles.length, successPages);
} catch (Exception e) {
log.warn("[DocumentExtract] OCR 失败: {}", e.getMessage());

3
mateclaw-ui/src/composables/chat/useChat.ts
View File

@ -630,11 +630,12 @@ export function useChat(options: UseChatOptions): UseChatReturn {
assistantMessage.conversationId = conversationId
currentAssistantId.value = assistantMessage.id as string
// contentParts 已由 buildOutgoingParts 包含 file entries，不要重复合并 attachments
await stream.connect({
agentId,
message: content,
conversationId,
contentParts: [...contentParts, ...attachments],
contentParts,
})
} catch (e) {
error.value = e instanceof Error ? e : new Error(String(e))