Enhance table title handling and improve heading detection logic

- Added logic to set `table_title` from `caption` if it is under 160 characters and not already set.
- Updated `_blocks_from_content_list` to assign `table_title` based on `caption` length.
- Introduced new regex patterns for better detection of TOC entries and noise.
- Enhanced heading detection to differentiate between numbered instructions and actual headings.
- Added tests to verify that table captions are correctly assigned as titles and that numbered instructions are treated as body text.

backend/rag_cut/splitters/pdf_semantic.py

@@ -8,12 +8,20 @@ from dataclasses import dataclass, field
 from rag_cut.models import Block, BlockType, Chunk, SplitConfig
 from rag_cut.parsers.pdf.noise_filter import is_toc_noise_text, is_toc_title_text
 from rag_cut.renderer import collect_chunk_layout_meta, render_blocks
+from rag_cut.splitters.heading_splitter import (
+    PURE_NUMBERED_HEADING_RE,
+    TIME_LIKE_RE,
+    looks_like_false_heading_text,
+    looks_like_numbered_instruction,
+    normalize_heading_block,
+)
 
 CHUNK_STRATEGY = "heading_layout_multimodal"
 
 PAGE_NUMBER_RE = re.compile(r"^\s*(?:[-\u2013\u2014]?\s*)?\d{1,4}(?:\s*/\s*\d{1,4})?\s*$")
 NUMBERED_HEADING_RE = re.compile(
-    r"^\s*(?P<num>\d+(?:\.\d+)*)(?:\.|．)?\s*(?P<title>[A-Za-z0-9\u4e00-\u9fff][^\n]{1,120})\s*$"
+    r"^\s*(?P<num>\d+(?:\.\d+)*)(?:[.．、:：]\s*|\s+)"
+    r"(?P<title>[A-Za-z\u4e00-\u9fff][^\n]{1,120})\s*$"
 )
 LETTER_HEADING_RE = re.compile(
     r"^\s*(?P<letter>[A-Z])[\.)]\s+(?P<title>[A-Za-z0-9\u4e00-\u9fff][^\n]{1,100})\s*$"
@@ -132,7 +140,7 @@ def _is_noise(block: Block, running_headers: set[str]) -> bool:
     text = " ".join(_text(block).split())
     if block.type == BlockType.IMAGE:
         return _is_decorative_image(block)
-    if PAGE_NUMBER_RE.match(text):
+    if PAGE_NUMBER_RE.match(text) or TIME_LIKE_RE.match(text):
         return True
     if _is_toc_text(text):
         return True
@@ -146,19 +154,29 @@ def _heading_signal(block: Block, current_top_level: bool = False) -> _HeadingSi
     if not text or len(text) > 180 or _is_toc_text(text):
         return None
     # Keep procedural steps inside the parent section; do not open a new group.
-    if STEP_INSTRUCTION_RE.match(text):
+    if (
+        STEP_INSTRUCTION_RE.match(text)
+        or looks_like_numbered_instruction(text)
+        or looks_like_false_heading_text(text)
+    ):
         return None
 
+    pure_numbered = PURE_NUMBERED_HEADING_RE.match(text)
+
     if block.type == BlockType.HEADING:
         level = block.level or 1
         numbered = NUMBERED_HEADING_RE.match(text)
-        if numbered:
+        if pure_numbered:
+            level = pure_numbered.group(1).count(".") + 1
+        elif numbered:
             level = numbered.group("num").count(".") + 1
         elif LETTER_HEADING_RE.match(text):
             level = 2 if current_top_level else max(2, level)
         return _HeadingSignal(text=" ".join(text.split()), level=max(1, min(level, 6)))
 
     numbered = NUMBERED_HEADING_RE.match(text)
+    if pure_numbered:
+        return _HeadingSignal(text=" ".join(text.split()), level=pure_numbered.group(1).count(".") + 1)
     if numbered:
         return _HeadingSignal(text=" ".join(text.split()), level=numbered.group("num").count(".") + 1)
     if LETTER_HEADING_RE.match(text):
@@ -339,7 +357,7 @@ def _split_oversized_group(group: _Group, start_index: int, config: SplitConfig)
 
     def flush() -> None:
         nonlocal current, current_len, part_index
-        body = [b for b in current if b not in prefix]
+        body = [b for b in current if b.type != BlockType.HEADING]
         if not body:
             return
         part_group = _Group(group.heading_path, list(current))
@@ -362,7 +380,17 @@ def _split_oversized_group(group: _Group, start_index: int, config: SplitConfig)
     for block in group.blocks[len(prefix) :]:
         block_len = len(block.render()) + 2
         if current_len + block_len > config.max_chunk_size and len(current) > len(prefix):
-            flush()
+            context_len = sum(
+                len(item.render()) + 2
+                for item in current
+                if item.type != BlockType.HEADING
+            )
+            keep_atomic_context = (
+                block.type in {BlockType.IMAGE, BlockType.TABLE}
+                and context_len <= min(400, max(80, config.max_chunk_size // 2))
+            )
+            if not keep_atomic_context:
+                flush()
         current.append(block)
         current_len += block_len
     flush()
@@ -386,9 +414,10 @@ def _build_groups(blocks: list[Block]) -> list[_Group]:
                 _attach_heading_only_to_previous(groups, current)
         current = _Group()
 
-    for raw in blocks:
-        if _is_noise(raw, running_headers):
+    for source in blocks:
+        if _is_noise(source, running_headers):
             continue
+        raw = normalize_heading_block(source)
 
         signal = _heading_signal(raw, current_top_level=bool(heading_stack))
         if signal:
@@ -399,16 +428,11 @@ def _build_groups(blocks: list[Block]) -> list[_Group]:
             heading_stack.append((signal.level, signal.text, heading))
             path = [item[1] for item in heading_stack]
             current = _Group(heading_path=path)
-            for _, _, h_block in heading_stack:
-                current.blocks.append(_enrich_block(h_block, path, current.blocks))
+            current.blocks.append(_enrich_block(heading, path, current.blocks))
             continue
 
         path = [item[1] for item in heading_stack]
-        if not current.blocks and heading_stack:
-            current.heading_path = path
-            for _, _, h_block in heading_stack:
-                current.blocks.append(_enrich_block(h_block, path, current.blocks))
-        elif not current.heading_path:
+        if not current.heading_path:
             current.heading_path = path
 
         current.blocks.append(_enrich_block(raw, current.heading_path, current.blocks))