Skip to content

Formerly KrillinAI, VideoTranslation and Dubbing Tool for 120+ Languages. Open-source AI workspace for creators, powered by Codex. Video downloader, video translation, create videos, images, audios, and edits with Agents in one place. For your Social Media like YouTube, TikTok, Instagram, etc.

test(util): pin fence-then-walk order and the comma trim cutset

askalfd185f7ea431a
diff --git a/runtime/krillinai/pkg/util/json_extract_fence_test.go b/runtime/krillinai/pkg/util/json_extract_fence_test.go
new file mode 100644
index 00000000..3580ba8b
--- /dev/null
+++ b/runtime/krillinai/pkg/util/json_extract_fence_test.go
@@ -0,0 +1,72 @@
+package util
+
+import "testing"
+
+// TestExtractJSONObjectFenceAndWalkOrder 固定“先去 Markdown 围栏、再按候选向后扫描”这一顺序。
+// ExtractJSONObject 先调用 CleanMarkdownCodeBlock,再在结果上逐个取候选,
+// 因此围栏内外的示例对象都会先变成普通文本、再被候选扫描跳过。
+// 已有的“Markdown 代码块”用例只有围栏、没有示例对象,
+// “示例空对象被跳过”只有示例对象、没有围栏,两者的组合此前没有用例覆盖。
+// 修复前整段响应只去围栏就交给 encoding/json,两条都拿不到单个对象,能区分修复前后。
+func TestExtractJSONObjectFenceAndWalkOrder(t *testing.T) {
+ tests := []struct {
+ name string
+ response string
+ key string
+ want string
+ }{
+ {
+ name: "围栏内的示例对象被跳过",
+ response: "```json\n{} {\"align\":[{\"origin_part\":\"a\"}]}\n```",
+ key: "align",
+ want: `{"align":[{"origin_part":"a"}]}`,
+ },
+ {
+ name: "围栏之前的示例对象被跳过",
+ response: "示例:{\"note\":\"x\"}\n```json\n{\"align\":[{\"origin_part\":\"a\"}]}\n```",
+ key: "align",
+ want: `{"align":[{"origin_part":"a"}]}`,
+ },
+ {
+ name: "字符串里的围栏标记不影响候选边界",
+ response: "{\"note\":\"```json\"} {\"align\":[{\"origin_part\":\"a\"}]}",
+ key: "align",
+ want: `{"align":[{"origin_part":"a"}]}`,
+ },
+ }
+
+ for _, tt := range tests {
+ t.Run(tt.name, func(t *testing.T) {
+ if got := ExtractJSONObject(tt.response, tt.key); got != tt.want {
+ t.Errorf("ExtractJSONObject() = %q, want %q", got, tt.want)
+ }
+ })
+ }
+}
+
+// TestExtractJSONObjectNonASCIITrailingComma 固定 trimTrailingComma 的字符集边界。
+// 它只去掉 ASCII 逗号(cutset 为 " \t\r\n,"),中文全角逗号不在其中,
+// 因此带全角逗号的候选仍然是非法 JSON、会被 hasTopLevelKey 判为不命中并继续向后扫描,
+// 而不是被“修好”后当成回答。这保证容错只覆盖已知的尾随逗号形态,不会猜测其他写法。
+// 修复前整段响应直接交给 encoding/json,拿不到后面那个对象,能区分修复前后。
+func TestExtractJSONObjectNonASCIITrailingComma(t *testing.T) {
+ const response = `{"align":[1],} {"align":[{"origin_part":"a"}]}`
+ const want = `{"align":[{"origin_part":"a"}]}`
+
+ if got := ExtractJSONObject(response, "align"); got != want {
+ t.Errorf("ExtractJSONObject() = %q, want %q", got, want)
+ }
+}
+
+// TestExtractJSONObjectDuplicateTopLevelKey 固定顶层字段重复时仍算命中。
+// encoding/json 解到 map 时后出现的同名字段会覆盖前一个,但字段依然存在,
+// 因此这类响应属于“模型的真正回答”,应当就地返回,而不是跳过后落到原样返回。
+// 前面放一个不含目标字段的示例对象,使“接受该候选”和“跳到最后原样返回”的结果不同。
+func TestExtractJSONObjectDuplicateTopLevelKey(t *testing.T) {
+ const response = `{"note":"见下"} {"align":[1],"align":[2]}`
+ const want = `{"align":[1],"align":[2]}`
+
+ if got := ExtractJSONObject(response, "align"); got != want {
+ t.Errorf("ExtractJSONObject() = %q, want %q", got, want)
+ }
+}