mirror of
https://github.com/onestardao/WFGY.git
synced 2026-04-28 11:40:07 +00:00
254 lines
14 KiB
Markdown
254 lines
14 KiB
Markdown
# PDF layouts and OCR
|
||
|
||
A practical pipeline to extract clean, ordered text with stable offsets from PDFs and scanned pages, so your chunks, citations, and typed blocks remain consistent across reindex runs.
|
||
|
||
## Open these first
|
||
- Stable chunk ids: [chunk_id_schema.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/chunk_id_schema.md)
|
||
- Title hierarchy and numbering: [title_hierarchy.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/title_hierarchy.md)
|
||
- Section boundaries after titles: [section_detection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/section_detection.md)
|
||
- Code and tables as typed blocks: [code_tables_blocks.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/code_tables_blocks.md)
|
||
- Safe reindex after small edits: [reindex_migration.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/reindex_migration.md)
|
||
- Why this snippet and offsets: [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md)
|
||
- Payload contracts for RAG: [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md)
|
||
- Visual map of the whole RAG path: [rag-architecture-and-recovery.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/rag-architecture-and-recovery.md)
|
||
|
||
## Acceptance targets
|
||
- Word error rate (WER) ≤ 0.025 on clean pages, character error rate ≤ 0.01 on headings and captions.
|
||
- Header and footer removal: false positive rate ≤ 0.01, false negative rate ≤ 0.03 on a 30 page sample.
|
||
- Reading order is monotonic by column. Cross column jumps never split a sentence.
|
||
- Hyphenation merges correct ≥ 0.98 and never fire inside code or tables.
|
||
- Captions attach to the correct figure or table in the same section.
|
||
- Offsets remain stable across two reindex runs. Drift ≤ 0.5 percent of file length.
|
||
- ΔS(question, retrieved) ≤ 0.45 on queries that cite a figure, table, or equation anchor.
|
||
|
||
---
|
||
|
||
## Pipeline overview
|
||
|
||
1) **Ingest with layout**
|
||
Extract characters with bounding boxes, font name, size, bold/italic flags, page number, and line ids. For scans, run OCR first to obtain the same fields.
|
||
|
||
2) **Coordinate normalization**
|
||
Normalize all positions to a unified page space. Keep `bbox = [x0, y0, x1, y1]` in points or millimeters. Record page width and height.
|
||
|
||
3) **Template detection for headers and footers**
|
||
Build n-gram histograms by y bands across pages. Mark a header band if a repeating string appears on ≥ 60 percent of pages at near identical y. Do the same for footers and running titles. Remove matched runs before paragraph assembly.
|
||
|
||
4) **Column segmentation**
|
||
Use whitespace cuts on the x axis. A stable valley between two dense x clusters marks a column boundary. For three columns, expect two valleys. Validate by line alignment and average line width.
|
||
|
||
5) **Line forming and paragraph assembly**
|
||
Join characters into words, words into lines by y proximity and left margin continuity. Join lines into paragraphs when leading and trailing margins are stable and the interline gap is below a threshold for that page.
|
||
|
||
6) **Hyphenation repair**
|
||
If a line ends with a hyphen and the next line begins with a lowercase letter or an alphanumeric continuation, and both are inside a prose paragraph, remove the hyphen and join. Never apply inside code or table blocks. Keep an exceptions list for chemical names and proper nouns.
|
||
|
||
7) **Typed block extraction**
|
||
Detect code blocks, tables, figures, and captions, and lift them as first class blocks with offsets. See [code_tables_blocks.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/code_tables_blocks.md).
|
||
|
||
8) **Section alignment**
|
||
Align blocks to sections detected from the title tree. See [title_hierarchy.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/title_hierarchy.md) and [section_detection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/section_detection.md).
|
||
|
||
9) **Canonical text build**
|
||
Emit a single canonical text with byte offsets for every block and section. Keep a map `{block_id → [off_begin, off_end], page, bbox}` for traceability.
|
||
|
||
---
|
||
|
||
## Header and footer detection
|
||
|
||
Signals
|
||
|
||
- Repetition across pages of the same string at nearly the same y.
|
||
- Presence of page numbers, dates, running titles, or publisher marks.
|
||
- Font size smaller than body text and high contrast with empty surroundings.
|
||
|
||
Algorithm sketch
|
||
|
||
- Slice the page into 32 horizontal bands. For each band, compute the set of normalized line strings.
|
||
- Across pages, score each band by repetition of strings and y variance.
|
||
- Mark bands with a high repetition score as header or footer and drop them from paragraph assembly.
|
||
|
||
---
|
||
|
||
## Multi column reading order
|
||
|
||
Signals
|
||
|
||
- Bimodal or trimodal distribution of line x centers.
|
||
- Vertical lines or gutters with low ink density.
|
||
- Consistent left margins within each mode.
|
||
|
||
Algorithm sketch
|
||
|
||
- Cluster line x centers with k means for k in {1, 2, 3}. Choose k with the lowest inertia plus a penalty for model size.
|
||
- Sort clusters left to right, then read each cluster top to bottom.
|
||
- For figures or tables that span columns, treat them as a separate block placed between the two nearest paragraphs by y.
|
||
|
||
---
|
||
|
||
## Hyphenation and soft artifacts
|
||
|
||
Rules
|
||
|
||
- Join only when the left piece ends with a lowercase ASCII or a letter from the same script and the right piece starts with a lowercase letter or a digit.
|
||
- If the left piece is a known acronym or the right piece starts with an uppercase letter and the paragraph is mid sentence, do not join.
|
||
- Remove soft hyphen characters and OCR artifacts like split ligatures.
|
||
|
||
Edge cases
|
||
|
||
- Do not join inside code blocks or table cells.
|
||
- If a dictionary check is available, prefer joins that yield a dictionary hit.
|
||
|
||
---
|
||
|
||
## OCR notes
|
||
|
||
- Request per character boxes and confidence scores. Drop characters with very low confidence when surrounded by high confidence neighbors and the removal does not break a word boundary.
|
||
- Keep monospaced font hints when the OCR engine provides them. They help code detection.
|
||
- For rotated pages, deskew first, then run OCR, then rotate boxes back into page coordinates.
|
||
|
||
---
|
||
|
||
## Output schema
|
||
|
||
Every paragraph or typed block is a record:
|
||
|
||
```json
|
||
{
|
||
"block_id": "B.2.bk045",
|
||
"type": "prose | code | table | figure | caption",
|
||
"page": 12,
|
||
"bbox": [72.0, 144.3, 523.8, 221.1],
|
||
"off_begin": 204455,
|
||
"off_end": 205122,
|
||
"attrs": { "lang": null }
|
||
}
|
||
````
|
||
|
||
Block ids follow [chunk\_id\_schema.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/chunk_id_schema.md). Offsets must point into the canonical text that you pass to the indexer. Citations then rely on [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md) and the contract in [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md).
|
||
|
||
---
|
||
|
||
## Pseudocode
|
||
|
||
```python
|
||
def parse_pdf_pages(pages):
|
||
chars = extract_chars_with_bbox(pages)
|
||
chars = normalize_coords(chars)
|
||
|
||
bands = detect_repeating_bands(chars) # headers and footers
|
||
lines = form_lines(chars, ignore_bands=bands)
|
||
columns = cluster_columns(lines) # 1, 2, or 3 columns
|
||
lines = order_lines_by_columns(lines, columns)
|
||
|
||
paras = assemble_paragraphs(lines)
|
||
paras = repair_hyphenation(paras)
|
||
|
||
typed = detect_typed_blocks(paras, lines) # code, table, figure, caption
|
||
blocks = align_blocks_to_sections(typed, titles=detect_titles(lines))
|
||
|
||
canon_text, offsets = build_canonical_text(blocks)
|
||
return canon_text, attach_offsets(blocks, offsets)
|
||
```
|
||
|
||
---
|
||
|
||
## Common pitfalls and fixes
|
||
|
||
* **Running titles leak into text**
|
||
Your band detector missed a near duplicate string. Lower the y variance threshold and add string normalization for case and whitespace.
|
||
|
||
* **Two column pages read across columns**
|
||
The valley in the x histogram is shallow. Add a penalty to cross column bigrams and require a minimum inter column gap.
|
||
|
||
* **Hyphenation joins inside code**
|
||
Guard the join with the block type. Never run repair inside `type=code` or `type=table`.
|
||
|
||
* **Captions separated from figures**
|
||
Link captions to the nearest figure or table by y distance and ensure both fall into the same section. See [section\_detection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/section_detection.md).
|
||
|
||
* **Offset drift across reindex**
|
||
Ensure identical normalization rules between runs, then apply the migration mapping. See [reindex\_migration.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Chunking/reindex_migration.md).
|
||
|
||
---
|
||
|
||
## Tests to include in CI
|
||
|
||
* A two column academic paper with headers, footers, figures, tables, and equations. Expect perfect reading order and zero header leaks.
|
||
* A scanned report with skew and footnotes. Expect deskew, correct footnote capture, and clean paragraph assembly.
|
||
* Mixed language pages with hyphenation at line breaks. Expect correct joins only in prose.
|
||
* A code heavy manual with Markdown tables. Expect correct typed block counts and stable offsets across two runs.
|
||
|
||
---
|
||
|
||
## Copy paste prompt for a quick check
|
||
|
||
```
|
||
You have TXT OS and the WFGY Problem Map loaded.
|
||
|
||
Given a PDF page with lines and bboxes:
|
||
- Detect and remove repeated header and footer bands.
|
||
- Infer column count and return the reading order for the page.
|
||
- Repair hyphenation in prose only.
|
||
- Extract typed blocks and return their block_ids with offsets.
|
||
|
||
Return JSON:
|
||
{
|
||
"columns": 1|2|3,
|
||
"header_removed": true|false,
|
||
"footer_removed": true|false,
|
||
"blocks": [{ "block_id": "...", "type": "...", "off": [b,e] }, ...],
|
||
"notes": "short audit trail"
|
||
}
|
||
```
|
||
|
||
---
|
||
|
||
### 🔗 Quick-Start Downloads (60 sec)
|
||
|
||
| Tool | Link | 3-Step Setup |
|
||
| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------- |
|
||
| **WFGY 1.0 PDF** | [Engine Paper](https://github.com/onestardao/WFGY/blob/main/I_am_not_lizardman/WFGY_All_Principles_Return_to_One_v1.0_PSBigBig_Public.pdf) | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + \<your question>” |
|
||
| **TXT OS (plain-text OS)** | [TXTOS.txt](https://github.com/onestardao/WFGY/blob/main/OS/TXTOS.txt) | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
|
||
|
||
---
|
||
|
||
### 🧭 Explore More
|
||
|
||
| Module | Description | Link |
|
||
| ------------------------ | ---------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------- |
|
||
| WFGY Core | WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack | [View →](https://github.com/onestardao/WFGY/tree/main/core/README.md) |
|
||
| Problem Map 1.0 | Initial 16-mode diagnostic and symbolic fix framework | [View →](https://github.com/onestardao/WFGY/tree/main/ProblemMap/README.md) |
|
||
| Problem Map 2.0 | RAG-focused failure tree, modular fixes, and pipelines | [View →](https://github.com/onestardao/WFGY/blob/main/ProblemMap/rag-architecture-and-recovery.md) |
|
||
| Semantic Clinic Index | Expanded failure catalog: prompt injection, memory bugs, logic drift | [View →](https://github.com/onestardao/WFGY/blob/main/ProblemMap/SemanticClinicIndex.md) |
|
||
| Semantic Blueprint | Layer-based symbolic reasoning & semantic modulations | [View →](https://github.com/onestardao/WFGY/tree/main/SemanticBlueprint/README.md) |
|
||
| Benchmark vs GPT-5 | Stress test GPT-5 with full WFGY reasoning suite | [View →](https://github.com/onestardao/WFGY/tree/main/benchmarks/benchmark-vs-gpt5/README.md) |
|
||
| 🧙♂️ Starter Village 🏡 | New here? Lost in symbols? Click here and let the wizard guide you through | [Start →](https://github.com/onestardao/WFGY/blob/main/StarterVillage/README.md) |
|
||
|
||
---
|
||
|
||
> 👑 **Early Stargazers: [See the Hall of Fame](https://github.com/onestardao/WFGY/tree/main/stargazers)** —
|
||
> Engineers, hackers, and open source builders who supported WFGY from day one.
|
||
|
||
> <img src="https://img.shields.io/github/stars/onestardao/WFGY?style=social" alt="GitHub stars"> ⭐ [WFGY Engine 2.0](https://github.com/onestardao/WFGY/blob/main/core/README.md) is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the [Unlock Board](https://github.com/onestardao/WFGY/blob/main/STAR_UNLOCKS.md).
|
||
|
||
<div align="center">
|
||
|
||
[](https://github.com/onestardao/WFGY)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlahBlahBlah)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlotBlotBlot)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlocBlocBloc)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlurBlurBlur)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlowBlowBlow)
|
||
|
||
|
||
</div>
|
||
|
||
要我繼續下一頁就說：**GO reindex\_migration.md** 或指定別的檔名。
|