pagemark 0.1.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,228 @@
1
+ # Byte-compiled / optimized / DLL files
2
+ __pycache__/
3
+ *.py[codz]
4
+ *$py.class
5
+
6
+ # C extensions
7
+ *.so
8
+
9
+ # Distribution / packaging
10
+ .Python
11
+ build/
12
+ develop-eggs/
13
+ dist/
14
+ downloads/
15
+ eggs/
16
+ .eggs/
17
+ lib/
18
+ lib64/
19
+ parts/
20
+ sdist/
21
+ var/
22
+ wheels/
23
+ share/python-wheels/
24
+ *.egg-info/
25
+ .installed.cfg
26
+ *.egg
27
+ MANIFEST
28
+
29
+ # PyInstaller
30
+ # Usually these files are written by a python script from a template
31
+ # before PyInstaller builds the exe, so as to inject date/other infos into it.
32
+ *.manifest
33
+ *.spec
34
+
35
+ # Installer logs
36
+ pip-log.txt
37
+ pip-delete-this-directory.txt
38
+
39
+ # Unit test / coverage reports
40
+ htmlcov/
41
+ .tox/
42
+ .nox/
43
+ .coverage
44
+ .coverage.*
45
+ .cache
46
+ nosetests.xml
47
+ coverage.xml
48
+ *.cover
49
+ *.py.cover
50
+ *.lcov
51
+ .hypothesis/
52
+ .pytest_cache/
53
+ cover/
54
+
55
+ # Translations
56
+ *.mo
57
+ *.pot
58
+
59
+ # Django stuff:
60
+ *.log
61
+ local_settings.py
62
+ db.sqlite3
63
+ db.sqlite3-journal
64
+
65
+ # Flask stuff:
66
+ instance/
67
+ .webassets-cache
68
+
69
+ # Scrapy stuff:
70
+ .scrapy
71
+
72
+ # Sphinx documentation
73
+ docs/_build/
74
+
75
+ # PyBuilder
76
+ .pybuilder/
77
+ target/
78
+
79
+ # Jupyter Notebook
80
+ .ipynb_checkpoints
81
+
82
+ # IPython
83
+ profile_default/
84
+ ipython_config.py
85
+
86
+ # pyenv
87
+ # For a library or package, you might want to ignore these files since the code is
88
+ # intended to run in multiple environments; otherwise, check them in:
89
+ # .python-version
90
+
91
+ # pipenv
92
+ # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
93
+ # However, in case of collaboration, if having platform-specific dependencies or dependencies
94
+ # having no cross-platform support, pipenv may install dependencies that don't work, or not
95
+ # install all needed dependencies.
96
+ # Pipfile.lock
97
+
98
+ # UV
99
+ # Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
100
+ # This is especially recommended for binary packages to ensure reproducibility, and is more
101
+ # commonly ignored for libraries.
102
+ # uv.lock
103
+
104
+ # poetry
105
+ # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
106
+ # This is especially recommended for binary packages to ensure reproducibility, and is more
107
+ # commonly ignored for libraries.
108
+ # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
109
+ # poetry.lock
110
+ # poetry.toml
111
+
112
+ # pdm
113
+ # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
114
+ # pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
115
+ # https://pdm-project.org/en/latest/usage/project/#working-with-version-control
116
+ # pdm.lock
117
+ # pdm.toml
118
+ .pdm-python
119
+ .pdm-build/
120
+
121
+ # pixi
122
+ # Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
123
+ # pixi.lock
124
+ # Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
125
+ # in the .venv directory. It is recommended not to include this directory in version control.
126
+ .pixi/*
127
+ !.pixi/config.toml
128
+
129
+ # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
130
+ __pypackages__/
131
+
132
+ # Celery stuff
133
+ celerybeat-schedule*
134
+ celerybeat.pid
135
+
136
+ # Redis
137
+ *.rdb
138
+ *.aof
139
+ *.pid
140
+
141
+ # RabbitMQ
142
+ mnesia/
143
+ rabbitmq/
144
+ rabbitmq-data/
145
+
146
+ # ActiveMQ
147
+ activemq-data/
148
+
149
+ # SageMath parsed files
150
+ *.sage.py
151
+
152
+ # Environments
153
+ .env
154
+ .envrc
155
+ .venv
156
+ env/
157
+ venv/
158
+ ENV/
159
+ env.bak/
160
+ venv.bak/
161
+
162
+ # Spyder project settings
163
+ .spyderproject
164
+ .spyproject
165
+
166
+ # Rope project settings
167
+ .ropeproject
168
+
169
+ # mkdocs documentation
170
+ /site
171
+
172
+ # mypy
173
+ .mypy_cache/
174
+ .dmypy.json
175
+ dmypy.json
176
+
177
+ # Pyre type checker
178
+ .pyre/
179
+
180
+ # pytype static type analyzer
181
+ .pytype/
182
+
183
+ # Cython debug symbols
184
+ cython_debug/
185
+
186
+ # PyCharm
187
+ # JetBrains specific template is maintained in a separate JetBrains.gitignore that can
188
+ # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
189
+ # and can be added to the global gitignore or merged into this file. For a more nuclear
190
+ # option (not recommended) you can uncomment the following to ignore the entire idea folder.
191
+ .idea/
192
+
193
+ # Abstra
194
+ # Abstra is an AI-powered process automation framework.
195
+ # Ignore directories containing user credentials, local state, and settings.
196
+ # Learn more at https://abstra.io/docs
197
+ .abstra/
198
+
199
+ # Visual Studio Code
200
+ # Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore
201
+ # that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
202
+ # and can be added to the global gitignore or merged into this file. However, if you prefer,
203
+ # you could uncomment the following to ignore the entire vscode folder
204
+ # .vscode/
205
+ # Temporary file for partial code execution
206
+ tempCodeRunnerFile.py
207
+
208
+ # Ruff stuff:
209
+ .ruff_cache/
210
+
211
+ # PyPI configuration file
212
+ .pypirc
213
+
214
+ # Marimo
215
+ marimo/_static/
216
+ marimo/_lsp/
217
+ __marimo__/
218
+
219
+ # Streamlit
220
+ .streamlit/secrets.toml
221
+ # Generated test fixtures (rebuilt on demand by tests/conftest.py)
222
+ tests/fixtures/*.pdf
223
+
224
+ # Local tooling and scratch files
225
+ .claude/
226
+ .mcp.json
227
+ plan.md
228
+ test.pdf
@@ -0,0 +1,47 @@
1
+ # Changelog
2
+
3
+ All notable changes to this project are documented here.
4
+
5
+ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
6
+ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
7
+
8
+ ## [Unreleased]
9
+
10
+ ## [0.1.0] - 2026-10-04
11
+
12
+ Initial release.
13
+
14
+ ### Added
15
+
16
+ - `Pagemark` client with `convert()` and `aconvert()`. The async version is the
17
+ implementation; the sync one wraps it in `asyncio.run()`. Pages run
18
+ concurrently behind a semaphore (default 4).
19
+ - Works against any OpenAI-compatible endpoint. No provider constants, base
20
+ URLs, or model defaults are baked in — the caller supplies `base_url`,
21
+ `model`, and `api_key`. Key resolution order: explicit argument,
22
+ `OPENAI_API_KEY`, then an error naming the `base_url`.
23
+ - Page rendering with pypdfium2 under a pixel budget, so `Page.dpi` reports the
24
+ effective DPI after downscaling rather than the requested one.
25
+ - Text-layer anchoring: the PDF text layer and image-object placeholders are
26
+ sent alongside the page image, middle-elided to roughly 6k characters.
27
+ - Five guardrails on every model response — repetition, truncation, coverage,
28
+ hallucination, and refusal.
29
+ - A four-step escalation ladder (normal, reseed, high DPI, fallback prompt) with
30
+ a per-step confidence ceiling, ending in a text-layer fallback. A page that
31
+ fails every step is recorded with warnings instead of aborting the document.
32
+ - Cross-page assembly: running headers and footers are stripped, paragraphs
33
+ split across a page break are rejoined, and hyphenated line breaks are healed.
34
+ - `ModelProfile` protocol with a self-registering profile registry and one
35
+ built-in profile, `generic`.
36
+ - Typed output: `Document`, `Page`, `Block`, `Asset`, and `TokenUsage` pydantic
37
+ models. Every page carries `confidence`, `warnings`, and `source`.
38
+ - Embedded PDF images extracted per page. Binary payloads live in a private
39
+ attribute, so `model_dump_json()` stays free of them.
40
+ - `Document.save(md_path, assets_dir=...)` writes the Markdown and the images.
41
+ - Page selection via `pages`, accepting `"3"`, `"1-5"`, `"1-5,9,12-"`, or a
42
+ `list[int]`. Out-of-range indices raise rather than being clipped silently.
43
+ - Optional CLI (`pagemark[cli]`) with `convert` and `doctor` commands.
44
+ - `py.typed` marker; the package type-checks under mypy strict.
45
+
46
+ [Unreleased]: https://github.com/uts58/pagemark/compare/v0.1.0...HEAD
47
+ [0.1.0]: https://github.com/uts58/pagemark/releases/tag/v0.1.0
pagemark-0.1.0/LICENSE ADDED
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 pagemark contributors
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
@@ -0,0 +1,226 @@
1
+ Metadata-Version: 2.5
2
+ Name: pagemark
3
+ Version: 0.1.0
4
+ Summary: Local-first PDF-to-Markdown converter using vision-language models
5
+ Project-URL: Homepage, https://github.com/uts58/pagemark
6
+ Project-URL: Repository, https://github.com/uts58/pagemark
7
+ Project-URL: Issues, https://github.com/uts58/pagemark/issues
8
+ Author: Utsha Saha
9
+ Maintainer: Glichin
10
+ License-Expression: MIT
11
+ License-File: LICENSE
12
+ Keywords: document-conversion,markdown,ocr,pdf,vision-language-model,vlm
13
+ Classifier: Development Status :: 3 - Alpha
14
+ Classifier: Intended Audience :: Developers
15
+ Classifier: Operating System :: OS Independent
16
+ Classifier: Programming Language :: Python :: 3
17
+ Classifier: Programming Language :: Python :: 3.10
18
+ Classifier: Programming Language :: Python :: 3.11
19
+ Classifier: Programming Language :: Python :: 3.12
20
+ Classifier: Programming Language :: Python :: 3.13
21
+ Classifier: Topic :: Text Processing :: Markup :: Markdown
22
+ Classifier: Typing :: Typed
23
+ Requires-Python: >=3.10
24
+ Requires-Dist: langchain-openai<2,>=1.0
25
+ Requires-Dist: markdown-it-py>=3
26
+ Requires-Dist: pillow>=10
27
+ Requires-Dist: pydantic>=2
28
+ Requires-Dist: pypdfium2>=4
29
+ Provides-Extra: all
30
+ Requires-Dist: datasets>=2; extra == 'all'
31
+ Requires-Dist: rapidfuzz>=3; extra == 'all'
32
+ Requires-Dist: rich>=13; extra == 'all'
33
+ Requires-Dist: typer>=0.9; extra == 'all'
34
+ Provides-Extra: cli
35
+ Requires-Dist: rich>=13; extra == 'cli'
36
+ Requires-Dist: typer>=0.9; extra == 'cli'
37
+ Provides-Extra: eval
38
+ Requires-Dist: datasets>=2; extra == 'eval'
39
+ Requires-Dist: rapidfuzz>=3; extra == 'eval'
40
+ Description-Content-Type: text/markdown
41
+
42
+ # pagemark
43
+
44
+ [![PyPI](https://img.shields.io/pypi/v/pagemark)](https://pypi.org/project/pagemark/)
45
+ [![Python](https://img.shields.io/pypi/pyversions/pagemark)](https://pypi.org/project/pagemark/)
46
+ [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://github.com/uts58/pagemark/blob/main/LICENSE)
47
+
48
+ pagemark converts PDFs to Markdown with vision-language models.
49
+
50
+ Use it with any OpenAI-compatible endpoint. You provide the `base_url`, `model`,
51
+ and `api_key`. pagemark renders each page, grounds the model with the PDF text
52
+ layer, and checks the result for common failures: repetition loops, truncation,
53
+ dropped columns, and hallucinated tables. If a check fails, it retries with a
54
+ stricter strategy and falls back to the text layer if the model still cannot
55
+ recover.
56
+
57
+ If one page fails, the document still finishes. pagemark records warnings for
58
+ that page instead of aborting the whole conversion.
59
+
60
+ ## Prerequisites
61
+
62
+ - Python 3.10+
63
+ - An OpenAI-compatible endpoint serving a vision-language model
64
+ (Ollama, vLLM, LM Studio, any hosted provider, etc.)
65
+
66
+ ## Install
67
+
68
+ ```bash
69
+ pip install pagemark # library only
70
+ pip install "pagemark[cli]" # with CLI
71
+ ```
72
+
73
+ ## Quick start
74
+
75
+ ```python
76
+ from pagemark import Pagemark
77
+
78
+ client = Pagemark(
79
+ base_url="http://localhost:11434/v1", # OpenAI-compatible API
80
+ api_key="your-key", # API key
81
+ )
82
+
83
+ doc = client.convert("paper.pdf", model="qwen3-vl:8b")
84
+
85
+ print(doc.markdown)
86
+ doc.save("out.md", assets_dir="out_assets/")
87
+ ```
88
+
89
+ `api_key` is optional. pagemark resolves it from the explicit argument first,
90
+ then from `OPENAI_API_KEY`. If neither exists, it raises an error that names the
91
+ `base_url`.
92
+
93
+ ### Async
94
+
95
+ `aconvert()` is the implementation behind `convert()`. `convert()` wraps it with
96
+ `asyncio.run()`. Pages run concurrently, bounded by a semaphore. The default
97
+ limit is 4.
98
+
99
+ ```python
100
+ from pagemark import Pagemark
101
+
102
+ client = Pagemark(
103
+ base_url="http://localhost:11434/v1", # OpenAI-compatible API
104
+ api_key="your-key", # API key
105
+ )
106
+
107
+ doc = await client.aconvert(
108
+ "paper.pdf",
109
+ model="qwen3-vl:8b",
110
+ concurrency=8,
111
+ )
112
+ ```
113
+
114
+ ### Page selection
115
+
116
+ Convert selected pages instead of the whole document:
117
+
118
+ ```python
119
+ doc = client.convert("paper.pdf", model="qwen3-vl:8b", pages="1-5,9,12-")
120
+ ```
121
+
122
+ `pages` accepts `"3"`, `"1-5"`, `"1-5,9,12-"`, or a `list[int]`. Out-of-range
123
+ indices raise an error instead of being silently clipped.
124
+
125
+ ### Bring your own model
126
+
127
+ If you need settings the constructor does not expose, pass a pre-configured
128
+ `BaseChatModel`. That covers custom timeouts, proxies, non-OpenAI LangChain
129
+ backends like `ChatAnthropic` or `ChatOllama`, callbacks, and similar options.
130
+
131
+ ```python
132
+ from langchain_openai import ChatOpenAI
133
+ from pagemark import Pagemark
134
+
135
+ my_model = ChatOpenAI(
136
+ model="...",
137
+ base_url="...",
138
+ api_key="...",
139
+ timeout=120,
140
+ max_retries=5,
141
+ )
142
+
143
+ client = Pagemark(chat_model=my_model)
144
+ doc = client.convert("paper.pdf")
145
+ ```
146
+
147
+ ### CLI
148
+
149
+ ```bash
150
+ pagemark convert paper.pdf -o paper.md \
151
+ --model qwen3-vl:8b \
152
+ --base-url http://localhost:11434/v1
153
+
154
+ pagemark convert paper.pdf -o paper.md --json paper.json \
155
+ --model qwen3-vl:8b \
156
+ --base-url http://localhost:11434/v1
157
+
158
+ pagemark doctor --base-url http://localhost:11434/v1 --model qwen3-vl:8b
159
+ ```
160
+
161
+ `doctor` checks whether the endpoint is reachable, whether the model is present,
162
+ and where the API key came from.
163
+
164
+ ## How it works
165
+
166
+ ```
167
+ pdf -> render + anchor -> prompt (profile) -> ChatOpenAI -> guardrails -> parse -> assemble -> Document
168
+ pypdfium2 per-model langchain retry ladder md-it stitching md + json
169
+ ```
170
+
171
+ 1. Render each page to a PNG with pypdfium2, downscaled to the profile's pixel
172
+ budget.
173
+ 2. Extract the PDF text layer as an anchor for the model. Long pages are
174
+ middle-elided to about 6k characters.
175
+ 3. Send the image and anchor text to the VLM.
176
+ 4. Check every response with five guardrails.
177
+ 5. Retry with stricter strategies when a guardrail fails.
178
+ 6. Assemble the pages into a `Document` with typed blocks, then stitch across
179
+ pages by stripping running headers and footers, joining split paragraphs, and
180
+ dehyphenating text.
181
+
182
+ ### Guardrails
183
+
184
+ | Check | What it catches |
185
+ |---|---|
186
+ | Repetition | n-gram loops, collapsed gzip ratio |
187
+ | Truncation | `finish_reason == "length"` |
188
+ | Coverage | anchor tokens missing from output; catches silently dropped columns |
189
+ | Hallucination | high rate of output tokens absent from anchor (born-digital pages) |
190
+ | Refusal | "I'm sorry", "I cannot", near-empty output on a non-blank page |
191
+
192
+ ### Escalation ladder
193
+
194
+ When any guardrail fails, pagemark retries with more aggressive strategies:
195
+
196
+ | Step | Strategy | Confidence ceiling |
197
+ |---|---|---|
198
+ | 1 | Normal render, standard prompt | 0.95 |
199
+ | 2 | Reseed: `temperature=0.3`, `seed=42` | 0.85 |
200
+ | 3 | High DPI: DPI x1.5 (capped at 400), pixels x1.5 | 0.75 |
201
+ | 4 | Fallback prompt: simpler instruction | 0.60 |
202
+ | 5 | Text layer fallback | 0.30 |
203
+
204
+ Each page includes `source`, `confidence`, and `warnings`, so downstream
205
+ consumers can filter by quality.
206
+
207
+ ## Output
208
+
209
+ `convert()` returns a `Document` with:
210
+
211
+ - `markdown`: the full stitched Markdown output
212
+ - `pages`: list of `Page` objects, each with:
213
+ - `markdown`: per-page Markdown
214
+ - `blocks`: typed blocks (heading, paragraph, table, figure, code, list, formula, caption, footnote)
215
+ - `source`: `"vlm"` or `"text_layer"` (if the model couldn't handle the page)
216
+ - `confidence`: 0-1 score
217
+ - `warnings`: what went wrong, if anything
218
+ - `assets`: embedded images extracted from the PDF
219
+ - `model_dump_json()`: full structured output as JSON
220
+ - `save(path, assets_dir=...)`: writes the Markdown file and extracted images
221
+
222
+ ## License
223
+
224
+ MIT. Runtime dependencies are permissively licensed - mostly MIT/Apache-2.0/BSD,
225
+ with a few under MPL-2.0 (certifi, orjson) and PSF-2.0 (typing-extensions). No
226
+ GPL/LGPL/AGPL dependencies in the tree.
@@ -0,0 +1,185 @@
1
+ # pagemark
2
+
3
+ [![PyPI](https://img.shields.io/pypi/v/pagemark)](https://pypi.org/project/pagemark/)
4
+ [![Python](https://img.shields.io/pypi/pyversions/pagemark)](https://pypi.org/project/pagemark/)
5
+ [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://github.com/uts58/pagemark/blob/main/LICENSE)
6
+
7
+ pagemark converts PDFs to Markdown with vision-language models.
8
+
9
+ Use it with any OpenAI-compatible endpoint. You provide the `base_url`, `model`,
10
+ and `api_key`. pagemark renders each page, grounds the model with the PDF text
11
+ layer, and checks the result for common failures: repetition loops, truncation,
12
+ dropped columns, and hallucinated tables. If a check fails, it retries with a
13
+ stricter strategy and falls back to the text layer if the model still cannot
14
+ recover.
15
+
16
+ If one page fails, the document still finishes. pagemark records warnings for
17
+ that page instead of aborting the whole conversion.
18
+
19
+ ## Prerequisites
20
+
21
+ - Python 3.10+
22
+ - An OpenAI-compatible endpoint serving a vision-language model
23
+ (Ollama, vLLM, LM Studio, any hosted provider, etc.)
24
+
25
+ ## Install
26
+
27
+ ```bash
28
+ pip install pagemark # library only
29
+ pip install "pagemark[cli]" # with CLI
30
+ ```
31
+
32
+ ## Quick start
33
+
34
+ ```python
35
+ from pagemark import Pagemark
36
+
37
+ client = Pagemark(
38
+ base_url="http://localhost:11434/v1", # OpenAI-compatible API
39
+ api_key="your-key", # API key
40
+ )
41
+
42
+ doc = client.convert("paper.pdf", model="qwen3-vl:8b")
43
+
44
+ print(doc.markdown)
45
+ doc.save("out.md", assets_dir="out_assets/")
46
+ ```
47
+
48
+ `api_key` is optional. pagemark resolves it from the explicit argument first,
49
+ then from `OPENAI_API_KEY`. If neither exists, it raises an error that names the
50
+ `base_url`.
51
+
52
+ ### Async
53
+
54
+ `aconvert()` is the implementation behind `convert()`. `convert()` wraps it with
55
+ `asyncio.run()`. Pages run concurrently, bounded by a semaphore. The default
56
+ limit is 4.
57
+
58
+ ```python
59
+ from pagemark import Pagemark
60
+
61
+ client = Pagemark(
62
+ base_url="http://localhost:11434/v1", # OpenAI-compatible API
63
+ api_key="your-key", # API key
64
+ )
65
+
66
+ doc = await client.aconvert(
67
+ "paper.pdf",
68
+ model="qwen3-vl:8b",
69
+ concurrency=8,
70
+ )
71
+ ```
72
+
73
+ ### Page selection
74
+
75
+ Convert selected pages instead of the whole document:
76
+
77
+ ```python
78
+ doc = client.convert("paper.pdf", model="qwen3-vl:8b", pages="1-5,9,12-")
79
+ ```
80
+
81
+ `pages` accepts `"3"`, `"1-5"`, `"1-5,9,12-"`, or a `list[int]`. Out-of-range
82
+ indices raise an error instead of being silently clipped.
83
+
84
+ ### Bring your own model
85
+
86
+ If you need settings the constructor does not expose, pass a pre-configured
87
+ `BaseChatModel`. That covers custom timeouts, proxies, non-OpenAI LangChain
88
+ backends like `ChatAnthropic` or `ChatOllama`, callbacks, and similar options.
89
+
90
+ ```python
91
+ from langchain_openai import ChatOpenAI
92
+ from pagemark import Pagemark
93
+
94
+ my_model = ChatOpenAI(
95
+ model="...",
96
+ base_url="...",
97
+ api_key="...",
98
+ timeout=120,
99
+ max_retries=5,
100
+ )
101
+
102
+ client = Pagemark(chat_model=my_model)
103
+ doc = client.convert("paper.pdf")
104
+ ```
105
+
106
+ ### CLI
107
+
108
+ ```bash
109
+ pagemark convert paper.pdf -o paper.md \
110
+ --model qwen3-vl:8b \
111
+ --base-url http://localhost:11434/v1
112
+
113
+ pagemark convert paper.pdf -o paper.md --json paper.json \
114
+ --model qwen3-vl:8b \
115
+ --base-url http://localhost:11434/v1
116
+
117
+ pagemark doctor --base-url http://localhost:11434/v1 --model qwen3-vl:8b
118
+ ```
119
+
120
+ `doctor` checks whether the endpoint is reachable, whether the model is present,
121
+ and where the API key came from.
122
+
123
+ ## How it works
124
+
125
+ ```
126
+ pdf -> render + anchor -> prompt (profile) -> ChatOpenAI -> guardrails -> parse -> assemble -> Document
127
+ pypdfium2 per-model langchain retry ladder md-it stitching md + json
128
+ ```
129
+
130
+ 1. Render each page to a PNG with pypdfium2, downscaled to the profile's pixel
131
+ budget.
132
+ 2. Extract the PDF text layer as an anchor for the model. Long pages are
133
+ middle-elided to about 6k characters.
134
+ 3. Send the image and anchor text to the VLM.
135
+ 4. Check every response with five guardrails.
136
+ 5. Retry with stricter strategies when a guardrail fails.
137
+ 6. Assemble the pages into a `Document` with typed blocks, then stitch across
138
+ pages by stripping running headers and footers, joining split paragraphs, and
139
+ dehyphenating text.
140
+
141
+ ### Guardrails
142
+
143
+ | Check | What it catches |
144
+ |---|---|
145
+ | Repetition | n-gram loops, collapsed gzip ratio |
146
+ | Truncation | `finish_reason == "length"` |
147
+ | Coverage | anchor tokens missing from output; catches silently dropped columns |
148
+ | Hallucination | high rate of output tokens absent from anchor (born-digital pages) |
149
+ | Refusal | "I'm sorry", "I cannot", near-empty output on a non-blank page |
150
+
151
+ ### Escalation ladder
152
+
153
+ When any guardrail fails, pagemark retries with more aggressive strategies:
154
+
155
+ | Step | Strategy | Confidence ceiling |
156
+ |---|---|---|
157
+ | 1 | Normal render, standard prompt | 0.95 |
158
+ | 2 | Reseed: `temperature=0.3`, `seed=42` | 0.85 |
159
+ | 3 | High DPI: DPI x1.5 (capped at 400), pixels x1.5 | 0.75 |
160
+ | 4 | Fallback prompt: simpler instruction | 0.60 |
161
+ | 5 | Text layer fallback | 0.30 |
162
+
163
+ Each page includes `source`, `confidence`, and `warnings`, so downstream
164
+ consumers can filter by quality.
165
+
166
+ ## Output
167
+
168
+ `convert()` returns a `Document` with:
169
+
170
+ - `markdown`: the full stitched Markdown output
171
+ - `pages`: list of `Page` objects, each with:
172
+ - `markdown`: per-page Markdown
173
+ - `blocks`: typed blocks (heading, paragraph, table, figure, code, list, formula, caption, footnote)
174
+ - `source`: `"vlm"` or `"text_layer"` (if the model couldn't handle the page)
175
+ - `confidence`: 0-1 score
176
+ - `warnings`: what went wrong, if anything
177
+ - `assets`: embedded images extracted from the PDF
178
+ - `model_dump_json()`: full structured output as JSON
179
+ - `save(path, assets_dir=...)`: writes the Markdown file and extracted images
180
+
181
+ ## License
182
+
183
+ MIT. Runtime dependencies are permissively licensed - mostly MIT/Apache-2.0/BSD,
184
+ with a few under MPL-2.0 (certifi, orjson) and PSF-2.0 (typing-extensions). No
185
+ GPL/LGPL/AGPL dependencies in the tree.