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

feat(wiki): clean conversational preamble and code fences from transformation output

Browse Source
This commit is contained in:
matevip 2026-05-12 11:25:50 +08:00
parent f04ea58c8d
commit 62c9776b01
2 changed files with 51 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
.gitignore vendored
View File

@ -98,7 +98,6 @@ CLAUDE.md
# Codex CLI local artifacts
.codex/
.superpowers/
# Sync tooling local state (generated each run; report is intentionally tracked)
scripts/.*-sync-state.json

53
mateclaw-server/src/main/java/vip/mate/wiki/service/WikiTransformationExecutor.java
View File

@ -144,12 +144,16 @@ public class WikiTransformationExecutor {
ChatResponse resp = chatModel.call(new Prompt(List.of(
new SystemMessage(systemPrompt), new UserMessage(userPrompt))));
String output = (resp == null || resp.getResult() == null
String rawOutput = (resp == null || resp.getResult() == null
|| resp.getResult().getOutput() == null)
? null : resp.getResult().getOutput().getText();
if (output == null || output.isBlank()) {
if (rawOutput == null || rawOutput.isBlank()) {
throw new IllegalStateException("LLM returned empty output");
}
String output = cleanLlmOutput(rawOutput);
if (output.isBlank()) {
throw new IllegalStateException("LLM output was empty after cleanup");
}
// Honour a mid-flight cancel: the cancel endpoint flipped the run
// row to 'cancelled' while the LLM was still working. Drop the
// output and stop here rather than overwrite the cancelled state.
@ -252,6 +256,51 @@ public class WikiTransformationExecutor {
return (t == null || t.isBlank()) ? ("raw#" + raw.getId()) : t;
}
/** Recognises the conversational openers LLMs sometimes prepend even when
* the system prompt told them not to. Lines matching this pattern at the
* very start of the output are dropped. */
private static final java.util.regex.Pattern PREAMBLE_PATTERN =
java.util.regex.Pattern.compile(
"^\\s*(以下是|下面是|这是|根据您的要求|Sure(?:!|,)?|Of course[!,]?|Here(?:'s| is| are)|Certainly[!,]?|Got it[!,]?)[^\\n]*[:：][^\\n]*\\n+",
java.util.regex.Pattern.CASE_INSENSITIVE);
/**
* Normalise raw LLM output before persisting:
* <ul>
* <li>strip an outer markdown / language code fence (```markdown ... ``` or ``` ... ```)</li>
* <li>drop a conversational opener line ending with a colon</li>
* <li>trim whitespace</li>
* </ul>
* Conservative — only trims when the heuristic match is unambiguous,
* because over-trimming on a structured response would corrupt content.
*/
static String cleanLlmOutput(String text) {
if (text == null) return "";
String result = text.trim();
// Outer code fence ```lang? ... ```
if (result.startsWith("```")) {
int firstNewline = result.indexOf('\n');
if (firstNewline > 0 && result.endsWith("```")) {
String header = result.substring(3, firstNewline).trim();
// Only strip when the header is empty or looks like a language tag
// (markdown / md / json / yaml / text / plaintext) — never strip
// when the LLM used ``` as actual fenced content inside.
if (header.isEmpty() || header.matches("(?i)markdown|md|text|plaintext|json|yaml|yml|html?")) {
result = result.substring(firstNewline + 1, result.length() - 3).trim();
}
}
}
// Conversational opener line ending with a colon, followed by content.
java.util.regex.Matcher m = PREAMBLE_PATTERN.matcher(result);
if (m.find()) {
result = result.substring(m.end()).trim();
}
return result;
}
// ==================== Save-as-page ====================
/**