fix: optimize title to formula

app/services/glm_postprocess.py

@@ -34,13 +34,7 @@ def find_consecutive_repeat(s: str, min_unit_len: int = 10, min_repeats: int = 1
         return None
 
     pattern = re.compile(
-        r"(.{"
-        + str(min_unit_len)
-        + ","
-        + str(max_unit_len)
-        + r"}?)\1{"
-        + str(min_repeats - 1)
-        + ",}",
+        r"(.{" + str(min_unit_len) + "," + str(max_unit_len) + r"}?)\1{" + str(min_repeats - 1) + ",}",
         re.DOTALL,
     )
     match = pattern.search(s)
@@ -74,9 +68,7 @@ def clean_repeated_content(
         if count >= line_threshold and (count / total_lines) >= 0.8:
             for i, line in enumerate(lines):
                 if line == common:
-                    consecutive = sum(
-                        1 for j in range(i, min(i + 3, len(lines))) if lines[j] == common
-                    )
+                    consecutive = sum(1 for j in range(i, min(i + 3, len(lines))) if lines[j] == common)
                     if consecutive >= 3:
                         original_lines = content.split("\n")
                         non_empty_count = 0
@@ -113,6 +105,11 @@ def clean_formula_number(number_content: str) -> str:
 # GLMResultFormatter
 # ---------------------------------------------------------------------------
 
+# Matches content that consists *entirely* of a display-math block and nothing else.
+# Used to detect when a text/heading region was actually recognised as a formula by vLLM,
+# so we can correct the label before heading prefixes (## …) are applied.
+_PURE_DISPLAY_FORMULA_RE = re.compile(r"^\s*(?:\$\$[\s\S]+?\$\$|\\\[[\s\S]+?\\\])\s*$")
+
 # Label → canonical category mapping (mirrors GLM-OCR label_visualization_mapping)
 _LABEL_TO_CATEGORY: dict[str, str] = {
     # text
@@ -173,6 +170,19 @@ class GLMResultFormatter:
             item["native_label"] = item.get("native_label", item.get("label", "text"))
             item["label"] = self._map_label(item.get("label", "text"), item["native_label"])
 
+            # Label correction: layout may say "text" (or a heading like "paragraph_title")
+            # but vLLM recognised the content as a formula and returned $$…$$.  Without
+            # correction the heading prefix (##) would be prepended to the math block,
+            # producing broken output like "## $$ \mathbf{y}=… $$".
+            raw_content = (item.get("content") or "").strip()
+            if item["label"] == "text" and _PURE_DISPLAY_FORMULA_RE.match(raw_content):
+                logger.debug(
+                    "Label corrected text (native=%s) → formula: pure display-formula detected",
+                    item["native_label"],
+                )
+                item["label"] = "formula"
+                item["native_label"] = "display_formula"
+
             item["content"] = self._format_content(
                 item.get("content") or "",
                 item["label"],
@@ -262,9 +272,7 @@ class GLMResultFormatter:
                         content = content[: -len(e)].strip()
                     break
             if not content:
-                logger.warning(
-                    "Skipping formula region with empty content after stripping delimiters"
-                )
+                logger.warning("Skipping formula region with empty content after stripping delimiters")
                 return ""
             content = "$$\n" + content + "\n$$"
 
@@ -314,9 +322,7 @@ class GLMResultFormatter:
                     formula_content = items[i + 1].get("content", "")
                     merged_block = deepcopy(items[i + 1])
                     if formula_content.endswith("\n$$"):
-                        merged_block["content"] = (
-                            formula_content[:-3] + f" \\tag{{{num_clean}}}\n$$"
-                        )
+                        merged_block["content"] = formula_content[:-3] + f" \\tag{{{num_clean}}}\n$$"
                     merged.append(merged_block)
                     skip.add(i + 1)
                 continue  # always skip the formula_number block itself
@@ -328,9 +334,7 @@ class GLMResultFormatter:
                     formula_content = block.get("content", "")
                     merged_block = deepcopy(block)
                     if formula_content.endswith("\n$$"):
-                        merged_block["content"] = (
-                            formula_content[:-3] + f" \\tag{{{num_clean}}}\n$$"
-                        )
+                        merged_block["content"] = formula_content[:-3] + f" \\tag{{{num_clean}}}\n$$"
                     merged.append(merged_block)
                     skip.add(i + 1)
                     continue
@@ -390,9 +394,7 @@ class GLMResultFormatter:
             block["index"] = i
         return merged
 
-    def _format_bullet_points(
-        self, items: list[dict], left_align_threshold: float = 10.0
-    ) -> list[dict]:
+    def _format_bullet_points(self, items: list[dict], left_align_threshold: float = 10.0) -> list[dict]:
         """Add missing bullet prefix when a text block is sandwiched between two bullet items."""
         if len(items) < 3:
             return items
@@ -422,10 +424,7 @@ class GLMResultFormatter:
             if not (cur_bbox and prev_bbox and nxt_bbox):
                 continue
 
-            if (
-                abs(cur_bbox[0] - prev_bbox[0]) <= left_align_threshold
-                and abs(cur_bbox[0] - nxt_bbox[0]) <= left_align_threshold
-            ):
+            if abs(cur_bbox[0] - prev_bbox[0]) <= left_align_threshold and abs(cur_bbox[0] - nxt_bbox[0]) <= left_align_threshold:
                 cur["content"] = "- " + cur_content
 
         return items