opensource/mateclaw
2
0
Fork 0
mirror of https://gitee.com/mateos/mateclaw.git synced 2026-09-15 20:08:18 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix(conversation): render image/video/audio attachment markers for LLM (#66)

Browse Source
This commit is contained in:
matevip 2026-05-08 15:03:42 +08:00
parent d87ed0365a
commit f3b1cba6dc
1 changed files with 31 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
mateclaw-server/src/main/java/vip/mate/workspace/conversation/ConversationService.java
View File

@ -556,6 +556,7 @@ public class ConversationService {
case "text" -> appendSegment(text, part.getText()); case "text" -> appendSegment(text, part.getText());
case "thinking", "tool_call", "parse_error" -> { /* skip — frontend reads these from contentParts directly */ } case "thinking", "tool_call", "parse_error" -> { /* skip — frontend reads these from contentParts directly */ }
case "file" -> appendSegment(text, renderFilePart(part)); case "file" -> appendSegment(text, renderFilePart(part));
case "image", "video", "audio", "model3d" -> appendSegment(text, renderMediaPart(part));
default -> appendSegment(text, part.getText()); default -> appendSegment(text, part.getText());
} }
} }
@ -614,6 +615,36 @@ public class ConversationService {
return "[附件] " + name + "（路径: " + path + "）"; return "[附件] " + name + "（路径: " + path + "）";
} }
/**
* Render an image/video/audio/3D-model content part for the LLM prompt.
* <p>
* Without this marker, media parts are invisible in the rendered text — the LLM
* sees only the user's accompanying text and has no idea an attachment was sent.
* That fails closed when the multimodal Media injection in {@code BaseAgent} is
* upstream-stripped (model heuristic claims vision but the actual provider drops
* the image), leaving the agent to ask "which image?" for an attachment the user
* already uploaded. The path lets file-reading tools ({@code read_file},
* {@code extract_document_text}, {@code detect_file_type}) work as a fallback.
*/
private String renderMediaPart(MessageContentPart part) {
String label = switch (part.getType()) {
case "image" -> "[图片]";
case "video" -> "[视频]";
case "audio" -> "[音频]";
case "model3d" -> "[3D 模型]";
default -> "[附件]";
};
String name = safe(part.getFileName());
if (name.isBlank()) {
name = "未命名";
}
String path = safe(part.getPath());
if (path.isBlank()) {
return label + " " + name;
}
return label + " " + name + "（路径: " + path + "）";
}
private void appendSegment(StringBuilder builder, String text) { private void appendSegment(StringBuilder builder, String text) {
String safeText = safe(text); String safeText = safe(text);
if (safeText.isBlank()) { if (safeText.isBlank()) {