* fix(fonts): CID-as-Unicode passthrough and subscript/superscript merging
Add smart CID-as-Unicode passthrough for Identity-H fonts without
ToUnicode maps. Uses /W array median CID heuristic to distinguish
Unicode-CID PDFs (Chromium-generated) from GID-based subsets.
Add merge_subscript_items() pass that merges small-font items (<75%
of dominant font size, ≤4 chars, tightly adjacent) into parent items.
Fixes chemical formulas (NH3, H2O, KClO3), footnote references, and
subscript notation (vf, Hfg, m3/kg) that were previously orphaned
as separate text items.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(subscripts): restrict merge to purely numeric text only
Tighten subscript merging to only merge items containing ASCII digits
(0-9). This avoids false positives with ordinal indicators (º), letter
subscripts (sol, vf), and small bullet characters (▶) that caused
table restructuring regressions.
Numeric-only keeps the primary wins: chemical formulas (NH3, H2O),
footnote references, and unit notation (m2, m3).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(subscripts): restrict merge to parent text ending with a letter
Only merge numeric subscripts when the parent item's text ends with
an alphabetic character. Prevents false merges like "33" + "1" in
fractions (33 1/3%), table credit numbers after spaces, and footnote
refs after punctuation (land.1 → land. 1). Chemical formulas (NH3,
H2O, KClO3) still merge correctly since parent ends with a letter.
Reduces pdf-eval regressions from 13 to 2 (both are correct reversions
of over-aggressive footnote merging from the prior commit).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Points to firecrawl/lopdf@edf6279 which fixes decompressed_content()
for streams without a /Filter entry. Our xobjects.rs workaround remains
as a defensive fallback.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
lopdf's decompressed_content() fails on Form XObjects without a /Filter
entry (uncompressed streams). This caused all body text to be lost in PDFs
generated by pdfrw and similar tools that wrap page content in uncompressed
Form XObjects (e.g. Cambridge University Press excerpts).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
lopdf's decompressed_content() fails with DictKey("Filter") when a
ToUnicode stream has no /Filter entry (uncompressed raw text). Now
falls back to the raw stream.content when decompression fails.
Fixes Identity-H fonts with valid uncompressed ToUnicode CMaps
producing empty text (e.g. neoenergia tabela PDF).
Type3 fonts render each glyph as a custom drawing/bitmap. Without a
ToUnicode CMap the character codes can't be mapped to Unicode, so
extracted text is garbage. Pages using only Type3 fonts are now
excluded from pages_with_text and added to pages_needing_ocr.
Fixes Korean/CJK PDFs using Type3 fonts (e.g. D2Coding) being
classified as TextBased when no usable text can be extracted.
Document titles, column headers, and newsletter mastheads that repeat
on every page were being fully stripped by strip_repeated_lines. Now
the first page's occurrence is preserved so the content appears once.
Fixes missing titles like "VOICE OF SOUTH MARION" in tax certificate
PDFs and column headers in IRS forms.
When every page uses fonts with unresolvable gid-encoded glyphs,
the extracted text is unreliable garbage. Suppress the markdown
output so consumers know to use OCR instead.
Only triggers for gid-encoded fonts specifically, not for other
OCR signals (Identity-H without ToUnicode, template images) where
the text layer may still be partially useful.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat(detect): flag Identity-H fonts without ToUnicode for OCR
Cyrillic (and other non-Latin) PDFs with Type0/Identity-H encoded fonts
and no ToUnicode CMap produce garbage text from direct extraction. Two
fixes:
1. Detector: new `page_has_identity_h_no_tounicode` check adds affected
pages to `pages_needing_ocr` regardless of PDF classification.
2. Extraction: extend garbage-text safety net to TextBased PDFs — when
extracted text is <50% alphanumeric, drop the markdown, set
`has_encoding_issues`, and flag all pages for OCR.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor: replace integration tests with synthetic unit tests
Remove PDF fixture dependencies from detector and lib tests. Use
in-memory lopdf documents to test Identity-H/ToUnicode detection logic.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
When rect-based, line-based, and heuristic detection all fail to find
tables on a page, try building a table directly from the layout
engine's column boundaries. Handles borderless tabular layouts like
exam/reference grids where columns are defined purely by text
alignment.
Includes header-row column refinement: when a detected column
contains multiple header items, it gets split at the gap between
them to recover the correct number of columns.
Guards against false positives: requires ≥4 columns, ≤40 rows,
>50% multi-column rows, short cells (avg ≤40 chars), no prose
content, no dominant single column, and no structural elements
(≥6 rects or ≥4 lines) on the page.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fonts using raw glyph ID names (gidNNNNN) in their Differences
encoding cannot be decoded to Unicode without the original font's
cmap table. Detect this pattern during font parsing and add
affected pages to pages_needing_ocr so downstream consumers
know to use OCR instead.
Fixes text extraction on PDFs like Tezukuri_Food-Menu.pdf where
the main body font (AcuminVariableConcept) uses gid-encoded
glyphs — even PyMuPDF and ODL fail on these.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Pages with tens of thousands of vector-drawing rects (e.g. architectural
plans with 36k+ rects) caused cluster_rects and sub-rect deduplication to
spend 10-50s in O(n²) loops. No real table has thousands of cell rects,
so once a union-find component exceeds 2000 elements we skip further
comparisons, and skip the O(n²) sub-rect dedup entirely for such pages.
Benchmarks on 8 slow production PDFs show 10-40x speedups on vector-heavy
files (12s→0.3s) with no regression on other file types. All 342 tests
pass.
Two changes to improve balance sheet / financial table detection:
1. split_side_by_side: Don't split when one side is text labels
and the other is numeric data at matching Y positions. This
prevents splitting a single label+number table into two
independent regions.
2. try_add_label_column: After detecting a numeric-only table,
look for unclaimed text items to the left at matching Y
positions and prepend them as column 0 (row labels).
Tested on IN_Annual_Report_2017 balance sheet which now produces
proper 3-column tables (Label|2016|2017) instead of separated
number tables and paragraph text.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When rect clusters have valid bounding boxes but insufficient grid
structure (e.g. 2x2 edges from outer borders), emit their bounding
box as a RectHintRegion so the heuristic detector can be scoped to
the table area. Requires reasonable dimensions (100-600pt height,
≤500pt width) and ≥6 text items inside the region.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Screenshot PDFs (e.g., Chrome "Save as PDF") embed images inside
tiling Pattern resources rather than as direct XObject images.
Now traverses Pattern resources during image detection, correctly
classifying these as Mixed with OCR recommended.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The third-pass change to extract embedded cmaps for TrueType fonts with
WinAnsiEncoding caused apostrophes and other characters to be dropped
in subsetted fonts where the cmap doesn't cover all glyphs. The declared
encoding is authoritative for these fonts. The OCR text extraction fix
(invisible Tr=3 text) doesn't need this cmap override.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
OCR text layers and some PDF producers emit trailing spaces on each
text item, which combine with gap-based joining to produce double
spaces ("Vice President"). Now collapses runs of 2+ spaces to single
space within lines, preserving leading indentation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Scanned PDFs with OCR text layers use rendering mode 3 (invisible text)
positioned behind page images. Previously we skipped all Tr=3 text.
Now for Mixed/template PDFs, if normal extraction produces garbage or
empty output, we retry with invisible text included. This unlocks text
from OCR-generated PDFs without requiring external OCR.
Also adds Windows Unicode BMP (3,1) subtable support to the TrueType
cmap fallback, and allows TrueType fonts with explicit encoding to
extract their embedded cmap (OCR fonts often lie about encoding).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Scanned PDFs with JBIG2/tiled image strips were misclassified as
TextBased because no individual image tile exceeded the template
threshold. Now checks aggregate image area per page (≥2M pixels).
Also adds is_garbage_text() check: if a Mixed/template PDF's extracted
text is predominantly non-alphanumeric (<50%), upgrade to Scanned so
callers use proper OCR instead of the garbage text layer.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Multi-item lines (titles, section headers) that span across detected
columns were being split into individual column buckets, corrupting
newspaper detection and reading order. Add identify_spanning_lines()
that groups items by Y-proximity and marks lines wider than 1.3x the
widest column that have no gap at a detected gutter boundary.
Uses column-aware thresholds and gutter-interval gap detection for
precision — only fires on pages with detected columns, preserves
original item order via boolean mask.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a heading wraps across multiple visual lines, the structure tree
creates separate heading nodes per line, producing broken output like
`## Snow` / `## Lake`. Extend merge_heading_lines to consider struct-tree
roles (H1-H6) in addition to font-size heuristics. A 20-word combined
limit prevents merging body text mis-tagged as headings.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: tagged PDF structure tree support for semantic markdown generation
Parse /StructTreeRoot from tagged PDFs and use semantic roles (H1-H6, P,
LI, BlockQuote, Code, Caption) to improve markdown output. Structure tree
headings add to font-size heuristics without suppressing them. Coverage
threshold (≥50%) ensures only properly tagged PDFs activate this path.
Phase 1: Parse structure tree with role maps, MCID collection, flattening
Phase 2: Capture MCIDs from BMC/BDC operators, tag TextItems
Phase 3: Structure-aware markdown generation in convert loop
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: accumulate consecutive code lines into single fenced block
Per-line code fencing produced broken markdown for multi-line code
blocks (separate ``` open/close per line). Unify struct-tree Code
role and font-based monospace detection into a single is_code_line
check with in_code_block state for proper accumulation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test: add tagged PDF fixture with Firecrawl docs content
Synthetic 7-page PDF with rich structure tree exercising H1, H2, H3,
P, Code, LI, Caption, TH, TD roles. Generated via fpdf2 script.
Integration test verifies struct tree parsing and code fence output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: remove python PDF generator script from repo
Keep the generated fixture PDF but don't track the generator script.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: handle malformed bare-name struct types in tagged PDFs
Some PDF generators (e.g. fpdf2) write /S Code instead of /S /Code
in structure elements. lopdf silently drops these objects since bare
tokens are invalid PDF syntax. Add a pre-processor that scans for
known bare struct type names and prepends / before loading.
Unifies path and memory loading through the same fix pipeline.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: update lopdf dependency to main branch
The firecrawl/zlib-checksum-encrypted branch was merged and deleted.
Point to main which includes all previously merged fixes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: switch lopdf to upstream repo pinned at 845cd3d
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Narrow annotation columns beside wide body columns were being
Y-interleaved with body text, producing garbled reading order.
Add sidebar detection with width ratio, line balance, and sparse
density guards to correctly trigger sequential column reading.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Use character-width ratios instead of font_size for Canva-style PDFs:
- Single-char prev: gap/prev.width < 1.25
- Multi→single: gap/avg_prev_char_width < 1.25
- Multi→multi: page-level threshold (gap/font_size)
Also adds second Canva detection path for per-character rendering
without embedded spaces (>50% single-char items).
Improves ebgt7isj04ophcq word accuracy from ~3% to 64%.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Canva-generated PDFs render text character-by-character with CSS-style
letter-spacing (~0.5-0.9× font_size). The hardcoded 0.10 threshold
caused every character to get a space inserted ("K a r i b i b").
Detect Canva pages via fix_letterspaced_items (≥50% items match "a b c"
pattern), compute an IQR-based threshold (median × 1.55) on the gap
distribution BEFORE space removal, then propagate per-page thresholds
through PageThresholds → group_into_lines_with_thresholds → TextLine
→ should_join_items.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Drop column-width padding from table output. Saves ~37% bytes on
table-heavy PDFs. Markdown renders identically — padding was purely
cosmetic. Optimizes token efficiency for AI agent consumers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Expose the pdf-inspector Rust library as a Python package using PyO3 + maturin.
Python users can now `pip install` and use `import pdf_inspector` for PDF
classification, text extraction, and markdown conversion with native Rust speed.
Adds:
- src/python.rs: PyO3 bindings (process_pdf, detect_pdf, extract_text, etc.)
- pyproject.toml: maturin build configuration
- pdf_inspector.pyi: type stubs for IDE support
- tests/test_python.py: 21 pytest tests covering all Python API functions
- examples/basic_usage.py: example script demonstrating all features
- Updated README with Python quick start and API reference
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Arabic PDFs using ToUnicode maps to presentation forms (U+FB50-FDFF,
U+FE70-FEFE) produced unreadable output. Apply NFKC normalization to
convert back to base Arabic (U+0600-06FF), then reverse character order
within items that had presentation forms (visual→logical order).
Improves Arabic PDF word overlap with Mistral from ~4.5% to ~78.6%.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add bimodal gap analysis to find_column_boundaries for PDFs with many
narrow columns (e.g. 24-column train schedules at 26pt spacing). When
a clear gap between within-column jitter and between-column spacing is
detected, switch to edge-based clustering with a lower threshold.
Also raise heuristic max columns from 15 to 25, and relax column
consistency validation for very wide tables (>15 cols).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Tables with very wide cells (e.g. 230+ chars from merged schedule columns)
caused massive whitespace padding in every row. Capping alignment at 40
characters reduces output size significantly (franklin: 77KB→24KB,
Closed-Business: 571KB→502KB) without affecting content.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Subset fonts number GIDs by document encounter order, not sorted order.
The sequential remap scrambled character mappings, losing 2,532+ CJK
characters. Now promotes TrueType cmap fallback when it has more entries
than the primary ToUnicode CMap. Also adds CJK scoring to score_text()
as defense-in-depth.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Points to firecrawl/lopdf branch that falls back to raw deflate when
zlib adler32 checksum is corrupted after RC4 decryption. Fixes 5
encrypted BrokerCheck PDFs that were misclassified as scanned.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Rect-guided tables:
- Add try_build_rect_guided_table() to build tables from rect cluster
X positions as column boundaries, bypassing heuristic detection
- Split merged multi-number TextItems (e.g. "10 11 12...31") into
individual day-column cells using column-advancing boundary assignment
- Interpolate missing column boundaries for holiday/non-work days
that lack colored rects
- Strip tilde-leader noise from cells (legend text bleeding)
- Filter legend text beyond table area via max-X threshold
- Pass cluster_rects through RectHintRegion for downstream use
Extractor improvements:
- Deduplicate clip-path and fill-path rects before using as table hints
- Add font width fallback for missing glyph metrics via average width
- Improve column detection scoring to prefer balanced gutters
- Handle side-by-side layouts with hint-region derived split points
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add 48 new tests (140 → 188 total) covering previously untested private
functions with synthetic data. No PDF fixtures needed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Images nested inside Form XObjects (e.g. HuttoISDWorkPerks.pdf) were
invisible to the detector. Rewrote analyze_page_images to recursively
follow Form XObjects. Also added pages_with_vector_text tracking so
PDFs with vector-outlined text (e.g. qsop_21-0055.pdf) are correctly
classified as IMAGE-BASED with OCR recommended.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The retry mechanism (excluding page-background rects) was firing for any
grid rejection, including "column 0 is completely empty". This caused a
regression on HTM_02-01_Part_A where the retry accepted a bad 12x9 grid
instead of falling through to heuristic detection.
Introduce GridResult enum so the caller only retries on FewNonEmptyRows
(the specific symptom of propagate_merged_cells collapse), not on
structural failures like empty columns or low fill ratio.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
PDFs with full-page background fills (e.g. 6_KE_Chart_5-5-17.pdf) had
their table grids rejected because propagate_merged_cells collapsed all
text into row 0. Add a retry mechanism: normal detection first, then
if it fails and page-origin background rects are present (>=95% width,
>=90% height, near origin), retry with those rects excluded from X-edge
extraction and propagate_merged_cells.
The retry path uses strict validation (50% non-empty rows, 40% content
density, max 200 chars per cell) to avoid false positives on pages where
paragraph text would be incorrectly captured in the grid.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pages using nested Form XObject chains (page → wrapper Form → text Form)
were returning 0 text items because extract_form_xobject_text() didn't
handle the Do operator. Add Do/q/Q/cm handling with recursive descent
(max depth 5) and proper CTM tracking through the nesting.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pages with images AND very few text operators (<10) are image pages
with overlay text (headers/footers), not real text pages. This fixes
image-heavy PDFs like newspaper ads being misclassified as TextBased.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The firecrawl/fix-cmap-optional-slash branch caused a regression on
the Italian PDF (garbled CMap-encoded text). The EPEVER fix was a
wrong diagnosis — the PDF extracts correctly on upstream main.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Count path construction/painting operators (m, l, c, h, f, S, re, etc.)
during the existing content stream scan. Pages where path ops vastly
outnumber text ops (>=1000 path ops AND >200x text ops) contain
vector-outlined glyphs that need OCR instead of text extraction.
This fixes EPEVER-DataSheet-XTRA-N-G3-Series-3.pdf page 2 which has
41,263 path ops but only 33 Tj ops — now correctly classified as Mixed
with page 2 flagged for OCR.
Also removes the check_page_decode_health phase (Phase 1.5) which
required parsing content streams into operations — the path-based
detection is both simpler and more effective.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Switches to a new lopdf branch that includes the CMap parser fix for
PDFs with uncompressed ToUnicode streams that omit the leading slash
on metadata keys (CIDSystemInfo, CMapName, CMapType).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
PDFs like fare1011.pdf draw table row stripes and borders using
m/l/h/f* filled paths instead of `re` operators. Extract axis-aligned
rectangles from these filled subpaths as a third-priority rect source
(after `re` rects and clip-path rects).
Key changes:
- Track completed subpaths in `h` handler for f/f* rect extraction
- Update W/W* handler to read from pending_subpaths when pending_lines
is empty (after h clears them)
- Add row-stripe fallback in detect_rects when clustering produces no
large clusters (non-overlapping row stripes), with ≥15 rects and
≥10 rows required to prevent decorative fill false positives
- Revert body-font upper bound to 1.05x (no longer needed at 1.10x)
fare1011.pdf: 12-col heuristic table → 14-col rect-based table with
all fare values correctly captured.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extract axis-aligned rectangles from W/W* clip operators in content
streams — many PDFs define table cells as clipping paths instead of
stroked rects. Add merged-cluster fallback in detect_tables_from_rects()
that merges all cluster rects when per-cluster detection fails or only
produces narrow false-positives (≤3 columns). Uses rect Y-edges for
rows and text X-clustering for columns.
Also updates lopdf to firecrawl fork (fix-leading-whitespace branch).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Thread a shared HashSet<u8> through scan_content_for_text_operators and
scan_xobjects_in_resources instead of returning a count and re-scanning
with a separate collect_unique_chars_from_content function.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
PDFs like wildfire_waiver.pdf render each line as a separate image XObject
with only bullet characters as actual text operators. Add image-dominance
heuristic (image_count > 10 && image_count > text_ops * 3) and unique text
character check (>= 5 unique non-whitespace chars) so these pages are no
longer misclassified as text-based.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Only skip text with rendering mode 3 (truly invisible). Removes
fill color tracking that was incorrectly hiding white-on-dark text
in presentations and brochures.
- Lower gap threshold from 40pt to 30pt to catch narrower inter-table gaps
- Constrain candidates to middle 60% of page X range
- Use center-position balance check instead of left-edge counting
- Add balanced-cluster dedup (50pt) to distinguish side-by-side tables
from single wide tables with multiple column gaps
- Improve filter_rects_to_band with 70% containment check to exclude
page-wide background stripes
Split pages with two independent table regions (e.g. D'Addario tension
chart) into separate bands before running rect/line/heuristic detection.
The split function finds X-position gaps ≥40pt, validates that <2% of
items cross the split point, and requires ≥30% balance on each side.
Also adds heuristic table detection fallback in compute_layout_complexity.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
