docvortex 0.4.4__tar.gz → 0.4.6__tar.gz

{docvortex-0.4.4/src/docvortex.egg-info → docvortex-0.4.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docvortex
-Version: 0.4.4
+Version: 0.4.6
 Summary: A fast, multi-format document parsing and conversion engine
 License-Expression: MIT
 Project-URL: Repository, https://github.com/myhloli/DocVortex
@@ -36,7 +36,7 @@ Requires-Dist: pylatexenc<3,>=2.10
 Requires-Dist: latex2mathml<4,>=3.81.0
 Requires-Dist: mathml2omml==0.0.2
 Requires-Dist: resvg_py<0.6,>=0.5.0
-Requires-Dist: reportlab
+Requires-Dist: reportlab[accel]
 Requires-Dist: ziamath<0.14,>=0.13
 Requires-Dist: magika<1.1.0,>=0.6.2
 Requires-Dist: opencv-python>=4.11.0.86
@@ -54,7 +54,7 @@ Dynamic: license-file
 
 <div align="center">
 
-<img src="https://raw.githubusercontent.com/myhloli/DocVortex/main/docs/images/docvortex-logo.jpg" alt="DocVortex logo" width="200">
+<img src="https://gcore.jsdelivr.net/gh/myhloli/DocVortex@main/docs/images/docvortex-logo.jpg" alt="DocVortex logo" width="200">
 
 # DocVortex
 
@@ -85,7 +85,7 @@ then exports the result in the formats your workflow needs.
 Native parsing works without an OCR or VLM inference service. Use DocVortex
 directly through its CLI or Python SDK, independently of MinerU.
 
-![DocVortex pipeline: native documents become a unified representation, then Markdown, HTML, LaTeX, DOCX, EPUB, PDF or structured JSON.](https://raw.githubusercontent.com/myhloli/DocVortex/main/docs/images/docvortex-overview.jpg)
+![DocVortex pipeline: native documents become a unified representation, then Markdown, HTML, LaTeX, DOCX, EPUB, PDF or structured JSON.](https://gcore.jsdelivr.net/gh/myhloli/DocVortex@main/docs/images/docvortex-overview.jpg)
 
 ## Quick start
 
@@ -183,7 +183,7 @@ reads metadata without parsing the document body.
 
 - **Text PDFs:** native parsing uses the document's existing text and structure. Scanned pages requiring OCR need an external OCR or inference service.
 - **PDF classification:** `docvortex classify report.pdf` returns `txt` or `ocr`. Classification is explicit and does not start inference; parsing does not automatically switch backends.
-- **PDF export:** PDF sources with page geometry default to block layout restoration, with selectable text and region images for tables/charts. Other sources and older results use semantic reflow. Use `--pdf-layout original|reflow` to select explicitly; fonts, line breaks and drawing instructions are not reproduced losslessly. See [PDF output layout](docs/USAGE.md#pdf-output-layout).
+- **PDF export:** PDF sources with page geometry default to block layout restoration, with selectable text and HTML-based tables; charts retain region images. Other sources and older results use semantic reflow. Use `--pdf-layout original|reflow` to select explicitly; fonts, line breaks and drawing instructions are not reproduced losslessly. See [PDF output layout](docs/USAGE.md#pdf-output-layout).
 
 ## Documentation
 

{docvortex-0.4.4 → docvortex-0.4.6}/README.md

@@ -1,6 +1,6 @@
 <div align="center">
 
-<img src="https://raw.githubusercontent.com/myhloli/DocVortex/main/docs/images/docvortex-logo.jpg" alt="DocVortex logo" width="200">
+<img src="https://gcore.jsdelivr.net/gh/myhloli/DocVortex@main/docs/images/docvortex-logo.jpg" alt="DocVortex logo" width="200">
 
 # DocVortex
 
@@ -31,7 +31,7 @@ then exports the result in the formats your workflow needs.
 Native parsing works without an OCR or VLM inference service. Use DocVortex
 directly through its CLI or Python SDK, independently of MinerU.
 
-![DocVortex pipeline: native documents become a unified representation, then Markdown, HTML, LaTeX, DOCX, EPUB, PDF or structured JSON.](https://raw.githubusercontent.com/myhloli/DocVortex/main/docs/images/docvortex-overview.jpg)
+![DocVortex pipeline: native documents become a unified representation, then Markdown, HTML, LaTeX, DOCX, EPUB, PDF or structured JSON.](https://gcore.jsdelivr.net/gh/myhloli/DocVortex@main/docs/images/docvortex-overview.jpg)
 
 ## Quick start
 
@@ -129,7 +129,7 @@ reads metadata without parsing the document body.
 
 - **Text PDFs:** native parsing uses the document's existing text and structure. Scanned pages requiring OCR need an external OCR or inference service.
 - **PDF classification:** `docvortex classify report.pdf` returns `txt` or `ocr`. Classification is explicit and does not start inference; parsing does not automatically switch backends.
-- **PDF export:** PDF sources with page geometry default to block layout restoration, with selectable text and region images for tables/charts. Other sources and older results use semantic reflow. Use `--pdf-layout original|reflow` to select explicitly; fonts, line breaks and drawing instructions are not reproduced losslessly. See [PDF output layout](docs/USAGE.md#pdf-output-layout).
+- **PDF export:** PDF sources with page geometry default to block layout restoration, with selectable text and HTML-based tables; charts retain region images. Other sources and older results use semantic reflow. Use `--pdf-layout original|reflow` to select explicitly; fonts, line breaks and drawing instructions are not reproduced losslessly. See [PDF output layout](docs/USAGE.md#pdf-output-layout).
 
 ## Documentation
 

{docvortex-0.4.4 → docvortex-0.4.6}/docs/JSON_PROTOCOL.md

@@ -113,8 +113,9 @@ when selecting a subset of an existing result.
 Pages may also contain `image_rotations`, such as `{"3":270}`: keys are source
 block indices encoded as strings, and values are the 0/90/180/270-degree angles
 used to turn region crops upright. The original-layout renderer reverses this
-crop transform when placing the image. This preserves rotated table images
-without duplicating image data or adding rotated-text reconstruction.
+crop transform when placing the image. Structured HTML tables use the same angle
+to restore the source orientation after layout in upright coordinates. No extra
+image data or schema fields are required.
 
 This uses the existing JSON extension mechanism; Model/Middle/bundle schema
 versions remain 2.0. Older documents without this extension remain readable.

{docvortex-0.4.4 → docvortex-0.4.6}/docs/USAGE.md

@@ -99,8 +99,22 @@ document and the original source file.
 PDF output defaults to `PdfLayout.AUTO`. Newly parsed PDF sources preserve their
 page dimensions, page boundaries, headers, footers, and block positions. Text is
 selectable and reflows inside each source block; original fonts and line breaks
-are approximate. Tables and charts prefer existing region images, and equations
-prefer vector rendering. Content absent from the parsed result cannot be restored.
+are approximate. Tables prefer parsed HTML with selectable text and vector borders;
+charts retain region images, and equations prefer vector rendering. Content absent
+from the parsed result cannot be restored.
+
+Original-layout tables use content-based column widths and compact cell padding.
+They first fit their source box, then may use safe space in the same column while
+preserving page margins and a 2 pt gap from other blocks. Existing spanning tables
+retain their span. Rotation is restored after layout in upright coordinates;
+tables do not move to another page. Captions and notes remain separate when their
+own coordinates are available, or share the parent box when child coordinates
+are missing. If 6 pt text still cannot fit, the table is scaled further and
+`pdf_layout_small_text` is reported. Small text alone never triggers an image
+fallback. Missing, invalid, or unrenderable HTML falls back to the region image,
+then text or a placeholder, with `pdf_table_fallback` and `pdf_table_source`
+diagnostics. `pdf_table_layout` records the source, rotation, effective font size,
+and original/drawing bounds. Reflow table layout is unchanged.
 
 ```python
 import docvortex
@@ -160,8 +174,9 @@ adjacent titles divide their gap at its midpoint. The original top edge is
 preferred, shifting upward when needed. Only a title that cannot fit its safe
 area shrinks, in 0.1 pt steps before the existing below-6-pt fallback. Script and
 inline-formula ink extents are included in title measurement. Index references
-and approximate parent groups do not participate; body text, captions, code and
-tables retain their original independent fitting results.
+and approximate parent groups do not participate; body text, captions and code
+retain their original independent fitting results. Structured tables are placed
+after titles and equations, respecting their final drawing bounds.
 
 `pdf_title_layout_expanded` identifies expanded title areas.
 `pdf_layout_font_exception` records local-body increases or insufficient-space

{docvortex-0.4.4 → docvortex-0.4.6}/pyproject.toml

@@ -27,7 +27,7 @@ dependencies = [
     "mammoth>=1.11.0,<2", "openpyxl>=3.1.5,<4", "olefile>=0.47,<1",
     "pylatexenc>=2.10,<3", "latex2mathml>=3.81.0,<4", "mathml2omml==0.0.2",
     "resvg_py>=0.5.0,<0.6",
-    "reportlab", "ziamath>=0.13,<0.14",
+    "reportlab[accel]", "ziamath>=0.13,<0.14",
     "magika>=0.6.2,<1.1.0", "opencv-python>=4.11.0.86",
 ]
 

{docvortex-0.4.4 → docvortex-0.4.6}/src/docvortex/analyzers/native/pdf/_script_geometry.py

@@ -208,42 +208,46 @@ def _cluster_baselines(
     median_loose_height = statistics.median(feature.loose_height for feature in anchors)
     tolerance = max(SCRIPT_BASELINE_ABSOLUTE_TOLERANCE, median_loose_height * SCRIPT_BASELINE_LOOSE_HEIGHT_RATIO)
     groups: list[list[int]] = []
+    group_origins: list[list[float]] = []
     for feature in sorted(anchors, key=lambda item: item.origin[1] if item.origin is not None else 0.0):
         baseline = feature.origin[1] if feature.origin is not None else 0.0
         if not groups:
             groups.append([feature.index])
+            group_origins.append([baseline])
             continue
-        previous_baseline = statistics.median(
-            features[index].origin[1] for index in groups[-1] if features[index].origin is not None
-        )
+        previous_baseline = _sorted_origin_median(group_origins[-1])
         if abs(baseline - previous_baseline) <= tolerance:
             groups[-1].append(feature.index)
+            group_origins[-1].append(baseline)
         else:
             groups.append([feature.index])
+            group_origins.append([baseline])
     return (
         [
             ScriptBaselineCluster(
-                baseline=statistics.median(features[index].origin[1] for index in group if features[index].origin is not None),
+                baseline=_sorted_origin_median(origins),
                 member_indices=tuple(group),
             )
-            for group in groups
+            for group, origins in zip(groups, group_origins, strict=True)
         ],
         tolerance,
     )
 
 
+def _sorted_origin_median(origins: list[float]) -> float:
+    """原始字符已按 origin 排序，直接使用与 statistics.median 相同的中位运算。"""
+    middle = len(origins) // 2
+    return origins[middle] if len(origins) % 2 else (origins[middle - 1] + origins[middle]) / 2
+
+
 def _cluster_tight_height(features: list[ScriptCharFeature], cluster: ScriptBaselineCluster) -> float:
     """返回基线簇的高分位 tight 高度。"""
-    return _quantile(
-        [features[index].tight_height for index in cluster.member_indices if features[index].tight_height > 0], 0.9
-    )
+    return _quantile([height for index in cluster.member_indices if (height := features[index].tight_height) > 0], 0.9)
 
 
 def _cluster_loose_height(features: list[ScriptCharFeature], cluster: ScriptBaselineCluster) -> float:
     """返回基线簇的高分位 loose 高度。"""
-    return _quantile(
-        [features[index].loose_height for index in cluster.member_indices if features[index].loose_height > 0], 0.9
-    )
+    return _quantile([height for index in cluster.member_indices if (height := features[index].loose_height) > 0], 0.9)
 
 
 def _cluster_tight_center(features: list[ScriptCharFeature], cluster: ScriptBaselineCluster) -> float:
@@ -283,29 +287,33 @@ def _cluster_has_consistent_displacement(
 def _choose_body_band(
     features: list[ScriptCharFeature],
     clusters: list[ScriptBaselineCluster],
+    heights: dict[ScriptBaselineCluster, tuple[float, float]] | None = None,
 ) -> tuple[ScriptBodyBand, ScriptBaselineCluster] | None:
     """在接近最高字形的基线簇中按字符数选择正文基线。"""
     if not clusters:
         return None
-    maximum_height = max(_cluster_tight_height(features, cluster) for cluster in clusters)
+    if heights is None:
+        heights = {
+            cluster: (_cluster_tight_height(features, cluster), _cluster_loose_height(features, cluster))
+            for cluster in clusters
+        }
+    maximum_height = max(heights[cluster][0] for cluster in clusters)
     comparable = [
-        cluster
-        for cluster in clusters
-        if _cluster_tight_height(features, cluster) >= maximum_height * SCRIPT_BODY_COMPARABLE_HEIGHT_RATIO
+        cluster for cluster in clusters if heights[cluster][0] >= maximum_height * SCRIPT_BODY_COMPARABLE_HEIGHT_RATIO
     ]
     body_cluster = max(
         comparable,
         key=lambda cluster: (
             len(cluster.member_indices),
-            _cluster_tight_height(features, cluster),
-            _cluster_loose_height(features, cluster),
+            heights[cluster][0],
+            heights[cluster][1],
         ),
     )
     return (
         ScriptBodyBand(
             baseline=body_cluster.baseline,
-            tight_height=_cluster_tight_height(features, body_cluster),
-            loose_height=_cluster_loose_height(features, body_cluster),
+            tight_height=heights[body_cluster][0],
+            loose_height=heights[body_cluster][1],
             member_indices=frozenset(body_cluster.member_indices),
         ),
         body_cluster,
@@ -508,6 +516,7 @@ def _recheck_weak_clusters_with_local_font_body(
     component_indices: list[int],
     body_band: ScriptBodyBand,
     cluster_roles: dict[ScriptBaselineCluster, ScriptRole],
+    heights: dict[ScriptBaselineCluster, tuple[float, float]],
 ) -> None:
     """用同字体局部正文撤销弱误判；不制造角标，不让撤销结果成为新参考。"""
     initial_roles = dict(cluster_roles)
@@ -543,7 +552,7 @@ def _recheck_weak_clusters_with_local_font_body(
             continue
         shift = cluster.baseline - reference.baseline
         strong_shift = abs(shift) >= height * SCRIPT_STRONG_SHIFT_RATIO
-        tight_ratio = _cluster_tight_height(features, cluster) / height
+        tight_ratio = heights[cluster][0] / height
         if abs(shift) < max(SCRIPT_ORIGIN_MIN_SHIFT_ABSOLUTE, height * SCRIPT_ORIGIN_MIN_SHIFT_RATIO) or (
             tight_ratio > SCRIPT_TIGHT_HEIGHT_RATIO and not (strong_shift and tight_ratio <= SCRIPT_STRONG_MAX_HEIGHT_RATIO)
         ):
@@ -559,7 +568,11 @@ def _assign_component(
 ) -> None:
     """在单个视觉组件内按 origin 基线簇和双 bbox 一致性分配角色。"""
     clusters, tolerance = _cluster_baselines(features, component_indices)
-    body_result = _choose_body_band(features, clusters)
+    # 基线簇在本次组件判定内不变，高度分位数可供正文选择与各角标规则共享。
+    heights = {
+        cluster: (_cluster_tight_height(features, cluster), _cluster_loose_height(features, cluster)) for cluster in clusters
+    }
+    body_result = _choose_body_band(features, clusters, heights)
     if body_result is None:
         return
     body_band, body_cluster = body_result
@@ -570,8 +583,8 @@ def _assign_component(
         if cluster is body_cluster:
             continue
         shift = cluster.baseline - body_band.baseline
-        tight_ratio = _cluster_tight_height(features, cluster) / body_band.tight_height
-        loose_ratio = _cluster_loose_height(features, cluster) / body_band.loose_height
+        tight_ratio = heights[cluster][0] / body_band.tight_height
+        loose_ratio = heights[cluster][1] / body_band.loose_height
         minimum_shift = max(SCRIPT_ORIGIN_MIN_SHIFT_ABSOLUTE, body_band.tight_height * SCRIPT_ORIGIN_MIN_SHIFT_RATIO)
         strong_shift = abs(shift) >= body_band.tight_height * SCRIPT_STRONG_SHIFT_RATIO
         if (
@@ -587,7 +600,7 @@ def _assign_component(
             cluster_roles[cluster] = "body"
         else:
             cluster_roles[cluster] = _script_role(shift)
-    _recheck_weak_clusters_with_local_font_body(features, font_keys, component_indices, body_band, cluster_roles)
+    _recheck_weak_clusters_with_local_font_body(features, font_keys, component_indices, body_band, cluster_roles, heights)
     for index in component_indices:
         feature = features[index]
         if (
@@ -607,6 +620,61 @@ def _assign_component(
         roles[index] = "body"
 
 
+def paired_script_roles(
+    chars: list[Char],
+    roles: list[ScriptRole],
+    tight_bboxes: dict[int, BBox],
+    origins: dict[int, tuple[float, float]],
+) -> dict[int, ScriptRole]:
+    """确认同一基字符右侧横向叠放的上下标，排除分式和远距邻列。"""
+
+    features = build_script_features(chars, tight_bboxes, origins, set())
+    visible = [feature for feature in features if feature.text.isprintable() and not feature.text.isspace()]
+    paired: dict[int, ScriptRole] = {}
+    for position, (base, first, second) in enumerate(zip(visible, visible[1:], visible[2:])):
+        if (
+            not all(feature.is_valid and feature.text.isalnum() for feature in (base, first, second))
+            or roles[base.index] != "body"
+            or {roles[first.index], roles[second.index]} != {"sup", "sub"}
+        ):
+            continue
+        assert base.tight_bbox is not None and base.origin is not None
+        assert first.tight_bbox is not None and first.origin is not None
+        assert second.tight_bbox is not None and second.origin is not None
+        scale = base.tight_height
+        if scale <= 0 or any(feature.tight_height > 0.8 * scale for feature in (first, second)):
+            continue
+        overlap = min(first.tight_bbox[2], second.tight_bbox[2]) - max(first.tight_bbox[0], second.tight_bbox[0])
+        minimum_width = min(first.tight_bbox[2] - first.tight_bbox[0], second.tight_bbox[2] - second.tight_bbox[0])
+        if overlap < 0.5 * minimum_width:
+            continue
+        if any(
+            not -0.15 * scale <= feature.tight_bbox[0] - base.tight_bbox[2] <= 0.5 * scale
+            or abs(feature.origin[1] - base.origin[1]) < 0.15 * scale
+            or abs(feature.origin[1] - base.origin[1]) > scale
+            for feature in (first, second)
+        ):
+            continue
+        if (first.origin[1] - base.origin[1]) * (second.origin[1] - base.origin[1]) >= 0:
+            continue
+        # 本补证据只接纳完整单字符角标，不能只给 grid 等多字母角标的首字母加样式。
+        following = visible[position + 3] if position + 3 < len(visible) else None
+        if (
+            following is not None
+            and following.is_valid
+            and following.text.isalnum()
+            and following.tight_bbox is not None
+            and following.origin is not None
+            and following.tight_height <= 0.8 * scale
+            and abs(following.origin[1] - second.origin[1]) <= 0.15 * scale
+            and -0.15 * scale <= following.tight_bbox[0] - second.tight_bbox[2] <= 0.5 * scale
+        ):
+            continue
+        paired[first.index] = roles[first.index]
+        paired[second.index] = roles[second.index]
+    return paired
+
+
 def classify_char_script_roles(
     chars: list[Char],
     *,

{docvortex-0.4.4 → docvortex-0.4.6}/src/docvortex/analyzers/native/pdf/_table_recovery/candidate.py

@@ -143,6 +143,14 @@ def _choose_cell_for_glyph_indexed(
         0,
         min(rows - 1, bisect_left(index.y_tracks, glyph.bbox[3]) - 1),
     )
+    if left_col == right_col and top_row == bottom_row:
+        # 只覆盖一个原子格时仍执行原面积/中心判定，省去候选集合与排序。
+        cell_index = index.owners[top_row][left_col]
+        bbox = specs[cell_index].bbox
+        if glyph_overlap_ratio(glyph, bbox) > 0:
+            return cell_index, False
+        center_x, center_y = bbox_center(glyph.bbox)
+        return (cell_index, False) if bbox[0] <= center_x <= bbox[2] and bbox[1] <= center_y <= bbox[3] else (None, False)
     candidate_indices = {
         index.owners[row][col] for row in range(top_row, bottom_row + 1) for col in range(left_col, right_col + 1)
     }

{docvortex-0.4.4 → docvortex-0.4.6}/src/docvortex/analyzers/native/pdf/_table_recovery/engine.py

@@ -19,6 +19,7 @@ from .contracts import (
 from .geometry import bbox_intersection, normalize_bbox
 from .sparse_hybrid import build_sparse_hybrid_candidates, diagnose_sparse_hybrid_candidate_builds
 from .sparse_multiline import build_sparse_multiline_candidates, diagnose_sparse_multiline_candidate_builds
+from .rule_band import build_rule_band_candidates
 from .text import build_native_table_text
 from .text_grid import build_text_candidates, diagnose_text_candidate_builds
 from .vector import MAX_PRIMITIVES_PER_TABLE, build_vector_candidates, diagnose_vector_candidate_builds
@@ -379,8 +380,11 @@ def _evaluate_native_pdf_table(
             )
         if _has_alias_affected_physical_blank_row(vector_attempts):
             sparse_hybrid_allowed = False
-    sparse_hybrid_candidates = build_sparse_hybrid_candidates(table_input, text) if sparse_hybrid_allowed else []
-    sparse_hybrid_selection = _select_candidate(sparse_hybrid_candidates)
+    rule_band_candidates = build_rule_band_candidates(table_input, text) if sparse_hybrid_allowed else []
+    sparse_hybrid_candidates = (
+        build_sparse_hybrid_candidates(table_input, text) if sparse_hybrid_allowed and not rule_band_candidates else []
+    )
+    sparse_hybrid_selection = _select_candidate([*rule_band_candidates, *sparse_hybrid_candidates])
     text_candidates = (
         build_text_candidates(table_input, text)
         if len(text.rows) >= 2 and vector_selection is None and sparse_hybrid_selection is None
@@ -396,6 +400,7 @@ def _evaluate_native_pdf_table(
             text_candidates = []
     existing_generated_candidates = [
         *vector_candidates,
+        *rule_band_candidates,
         *sparse_hybrid_candidates,
         *text_candidates,
     ]
@@ -466,7 +471,10 @@ def diagnose_native_pdf_table(table_input: NativeTableInput) -> dict[str, Any]:
         if evaluation.text is not None
         else ()
     )
-    sparse_multiline_attempts = (
+    rule_band_attempts: list[dict[str, Any]] = []
+    if evaluation.text is not None:
+        build_rule_band_candidates(table_input, evaluation.text, diagnostics=rule_band_attempts)
+    sparse_multiline_attempts = tuple(rule_band_attempts) + (
         diagnose_sparse_multiline_candidate_builds(
             table_input,
             evaluation.text,

docvortex-0.4.6/src/docvortex/analyzers/native/pdf/_table_recovery/rule_band.py

@@ -0,0 +1,135 @@
+"""以三线表物理表头边界恢复换行表头和同步重启的描述列。"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from .candidate import GridCellSpec, build_candidate
+from .contracts import NativeTableCandidate, NativeTableInput, NativeTableText
+from .geometry import normalize_angle, table_local_size
+from .sparse_common import _local_rules
+from .sparse_multiline import (
+    _build_column_hypothesis,
+    _has_overlapping_formula_rows,
+    _local_rectangles,
+    _row_occupancy,
+    _stable_gutters,
+)
+
+
+def build_rule_band_candidates(
+    table_input: NativeTableInput,
+    text: NativeTableText,
+    diagnostics: list[dict[str, Any]] | None = None,
+) -> list[NativeTableCandidate]:
+    """只在三条完整横线及无歧义列占用共同成立时恢复特殊三线表。"""
+
+    width, height = table_local_size(table_input.table_bbox, normalize_angle(table_input.angle))
+    rules = _local_rules(table_input, width, height)
+    tolerance = max(1.0, 0.15 * text.median_glyph_height)
+    horizontal = sorted(
+        {rule.coordinate for rule in rules if rule.orientation == "horizontal" and rule.end - rule.start >= 0.9 * width}
+    )
+    if (
+        len(horizontal) != 3
+        or horizontal[0] > tolerance
+        or height - horizontal[-1] > tolerance
+        or any(rule.orientation == "vertical" and tolerance < rule.coordinate < width - tolerance for rule in rules)
+    ):
+        return []
+    header_bottom = horizontal[1]
+    header = [row for row in text.rows if row.bbox[3] <= header_bottom + 0.5]
+    body = [row for row in text.rows if row.bbox[1] >= header_bottom - 0.5]
+    if not 1 <= len(header) <= 3 or len(body) < 3 or len(header) + len(body) != len(text.rows):
+        return []
+    if _has_overlapping_formula_rows(text, header_bottom):
+        return []
+    hypothesis = _build_column_hypothesis(text, width, rules, _local_rectangles(table_input, width, height), None)
+    if hypothesis is None:
+        return []
+    tracks = hypothesis.x_tracks
+    cols = len(tracks) - 1
+    glyphs = {glyph.glyph_id: glyph for glyph in text.glyphs}
+    occupancy = [_row_occupancy(row, glyphs, tracks) for row in body]
+    full = set(range(cols))
+    common = set.intersection(*occupancy)
+    prefix_cols = min(common) if common else 0
+    groups: list[tuple[int, int]] = []
+    if prefix_cols:
+        # 至少两个描述列同步重启，右侧逐行完整，才能区分 rowspan 与普通空格。
+        suffix = set(range(prefix_cols, cols))
+        starts = [index for index, occupied in enumerate(occupancy) if occupied == full]
+        ends = starts[1:] + [len(body)]
+        if (
+            prefix_cols < 2
+            or len(suffix) < 2
+            or common != suffix
+            or len(starts) < 3
+            or starts[0] != 0
+            or any(occupied not in (full, suffix) for occupied in occupancy)
+            or any(end - start < 2 for start, end in zip(starts, ends))
+        ):
+            return []
+        groups = list(zip(starts, ends))
+    elif any(occupied != full for occupied in occupancy):
+        return []
+    if not groups and len(header) == 1:
+        # 普通单行表头三线表仍走既有策略，避免无关候选和金标漂移。
+        return []
+    header_occupancy = [_row_occupancy(row, glyphs, tracks) for row in header]
+    if set.union(*header_occupancy) != full:
+        return []
+    if len(header) > 1 and header_occupancy[-1] != full:
+        # 多级表头可能有仅出现在上层的 rowspan 标签，不能误折叠成普通换行表头。
+        return []
+    margin = max(0.25, 0.05 * text.median_glyph_width)
+    if any(
+        token.bbox[0] + margin < boundary < token.bbox[2] - margin
+        for row in header
+        for token in row.tokens
+        for boundary in tracks[1:-1]
+    ):
+        return []
+    if not _stable_gutters(text, tracks, hypothesis.physical_boundaries, header_bottom, None):
+        return []
+    boundaries = [header_bottom]
+    for previous, current in zip(body, body[1:]):
+        if previous.bbox[3] > current.bbox[1] + 0.5:
+            return []
+        boundaries.append((previous.bbox[3] + current.bbox[1]) / 2.0)
+    boundaries.append(height)
+    specs = [GridCellSpec(0, col, 1, 1, (tracks[col], 0.0, tracks[col + 1], header_bottom)) for col in range(cols)]
+    spans = {start: end for start, end in groups}
+    for index in range(len(body)):
+        for col in range(cols):
+            if col < prefix_cols and index not in spans:
+                continue
+            end = spans[index] if col < prefix_cols else index + 1
+            specs.append(
+                GridCellSpec(index + 1, col, end - index, 1, (tracks[col], boundaries[index], tracks[col + 1], boundaries[end]))
+            )
+    record: dict[str, Any] = {}
+    candidate = build_candidate(
+        source="sparse_multiline",
+        rows=len(body) + 1,
+        cols=cols,
+        specs=tuple(specs),
+        text=text,
+        structure_support=1.0,
+        row_stability=1.0,
+        column_stability=1.0,
+        issues=("evidence=physical_header_band", f"wrapped_header_rows={len(header)}", f"descriptor_columns={prefix_cols}"),
+        use_grid_index=True,
+        diagnostics=record,
+    )
+    if diagnostics is not None:
+        diagnostics.append({"source": "sparse_multiline", "evidence": "physical_header_band", **record})
+    if (
+        candidate is None
+        or candidate.text_capture < 1.0
+        or candidate.order_consistency < 1.0
+        or candidate.score < 0.98
+        or record.get("ambiguous_glyph_ratio", 1.0) > 0.0
+    ):
+        return []
+    return [candidate]

{docvortex-0.4.4 → docvortex-0.4.6}/src/docvortex/analyzers/native/pdf/_table_recovery/text.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import statistics
+import re
 import unicodedata
 from dataclasses import dataclass
 
@@ -32,7 +33,7 @@ def _select_pending_glyphs(table_input: NativeTableInput) -> list[_PendingGlyph]
     table_bbox = normalize_bbox(table_input.table_bbox)
     if table_bbox is None:
         return []
-    selected: list[tuple[int, int, Char]] = []
+    selected: list[tuple[int, int, Char, BBox]] = []
     for fallback_index, char in enumerate(table_input.chars):
         char_bbox = normalize_bbox(char.get("bbox"))
         if char_bbox is None:
@@ -44,14 +45,14 @@ def _select_pending_glyphs(table_input: NativeTableInput) -> list[_PendingGlyph]
             source_index = int(char.get("char_idx", fallback_index))
         except (TypeError, ValueError):
             source_index = fallback_index
-        selected.append((source_index, fallback_index, char))
+        selected.append((source_index, fallback_index, char, char_bbox))
 
     selected.sort(key=lambda item: (item[0], item[1]))
     output: list[_PendingGlyph] = []
     angle = normalize_angle(table_input.angle)
     pending_space = False
     pending_break = False
-    for _source_sort, fallback_index, char in selected:
+    for source_index, _fallback_index, char, absolute_bbox in selected:
         raw_text = str(char.get("char") or "")
         if not raw_text:
             continue
@@ -66,9 +67,6 @@ def _select_pending_glyphs(table_input: NativeTableInput) -> list[_PendingGlyph]
         text = _normalize_table_text(raw_text)
         if not text or text.isspace():
             continue
-        absolute_bbox = normalize_bbox(char.get("bbox"))
-        if absolute_bbox is None:
-            continue
         local_bbox = page_bbox_to_table_local(
             absolute_bbox,
             table_bbox,
@@ -76,10 +74,6 @@ def _select_pending_glyphs(table_input: NativeTableInput) -> list[_PendingGlyph]
         )
         if local_bbox is None or local_bbox[3] - local_bbox[1] < 0.5:
             continue
-        try:
-            source_index = int(char.get("char_idx", fallback_index))
-        except (TypeError, ValueError):
-            source_index = fallback_index
         output.append(
             _PendingGlyph(
                 glyph_id=len(output),
@@ -153,7 +147,14 @@ def _assign_visual_rows(
             )
         else:
             rows[best_index].append(glyph)
-            row_bboxes[best_index] = bbox_union([row_bboxes[best_index], glyph.bbox])
+            # 两框合并保持原 min/max 顺序，避免为每个字符建立通用迭代器。
+            previous_bbox = row_bboxes[best_index]
+            row_bboxes[best_index] = (
+                min(previous_bbox[0], glyph.bbox[0]),
+                min(previous_bbox[1], glyph.bbox[1]),
+                max(previous_bbox[2], glyph.bbox[2]),
+                max(previous_bbox[3], glyph.bbox[3]),
+            )
             for row_index in range(best_index, len(row_bboxes)):
                 prefix_max_bottoms[row_index] = max(
                     prefix_max_bottoms[row_index - 1] if row_index > 0 else row_bboxes[row_index][3],
@@ -170,7 +171,11 @@ def _assign_visual_rows(
 def _contains_cjk(text: str) -> bool:
     """判断文本是否包含常见中日韩统一表意字符。"""
 
-    return any("\u3400" <= char <= "\u9fff" for char in text)
+    # 组行热路径通常只传入相邻两个字符，直接遍历避免逐次创建生成器。
+    for char in text:
+        if "\u3400" <= char <= "\u9fff":
+            return True
+    return False
 
 
 def _join_glyph_line(
@@ -208,6 +213,9 @@ def _looks_like_compact_unit_or_identifier(
     combined = previous_line.strip() + next_line.strip()
     if not combined or any(char.isspace() for char in combined):
         return False
+    # 带单位后缀的长英文标题仍有词界，不能仅因末尾 /% 就把相邻单词粘连。
+    if re.fullmatch(r"[A-Za-z]{4,}", previous_line.strip()) and re.match(r"[A-Za-z]{4,}(?:/|%)", next_line.strip()):
+        return False
     return any(char.isdigit() or char in "/_^%°µμ" for char in combined)
 
 
@@ -248,12 +256,14 @@ def _cell_row_separator(
 def _glyph_line_parts(
     glyphs: list[NativeTableGlyph] | list[_PendingGlyph],
     median_height: float,
+    *,
+    presorted: bool = False,
 ) -> list[tuple[str, int | None]]:
     """重建单行文本片段，并保留每个可见片段的原字符索引。"""
 
     if not glyphs:
         return []
-    ordered = sorted(glyphs, key=lambda glyph: (glyph.bbox[0], glyph.bbox[1], glyph.glyph_id))
+    ordered = glyphs if presorted else sorted(glyphs, key=lambda glyph: (glyph.bbox[0], glyph.bbox[1], glyph.glyph_id))
     parts: list[tuple[str, int | None]] = [(ordered[0].text, ordered[0].source_index)]
     previous = ordered[0]
     for glyph in ordered[1:]:
@@ -292,7 +302,8 @@ def _tokenize_row(
             groups[-1].append(glyph)
     tokens: list[NativeTableToken] = []
     for group in groups:
-        content = _join_glyph_line(group, median_height)
+        # group 是视觉行按同一排序键切出的连续片段，不必再次排序。
+        content = "".join(text for text, _source_index in _glyph_line_parts(group, median_height, presorted=True))
         if not content:
             continue
         tokens.append(