restruct-cv 0.1.0__tar.gz

restruct_cv-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Nyein Chan Soe
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

restruct_cv-0.1.0/PKG-INFO

@@ -0,0 +1,226 @@
+Metadata-Version: 2.4
+Name: restruct-cv
+Version: 0.1.0
+Summary: Turn messy PDF, DOCX, and scanned resumes into consistent, explainable, versioned JSON
+Keywords: resume,cv,parser,ats,applicant-tracking,ocr,pdf,docx
+Author: Nyein Chan Soe
+Author-email: Nyein Chan Soe <contact@nyeinchansoe.com>
+License-Expression: MIT
+License-File: LICENSE
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Text Processing
+Classifier: Topic :: Scientific/Engineering :: Information Analysis
+Requires-Dist: pillow>=11.0
+Requires-Dist: python-docx>=1.1
+Requires-Dist: pymupdf>=1.28.2
+Requires-Dist: sentence-transformers>=6.0.1
+Requires-Dist: transformers>=4.51.3,<5.14.0
+Requires-Python: >=3.11
+Project-URL: Homepage, https://github.com/Nye1nChanSoe/restruct-cv
+Project-URL: Repository, https://github.com/Nye1nChanSoe/restruct-cv
+Project-URL: Issues, https://github.com/Nye1nChanSoe/restruct-cv/issues
+Description-Content-Type: text/markdown
+
+# Restruct
+
+Restruct turns messy PDF, DOCX, and scanned resumes into consistent, explainable, versioned JSON.
+It combines OCR, layout geometry, NER, semantic similarity, and deterministic rules so every result
+can be traced back to the thing on the page that produced it.
+
+Built for applicant tracking systems, candidate search, analytics pipelines, talent platforms, and
+other systems that need trustworthy resume data.
+
+---
+
+## Install
+
+```bash
+uv sync                      # from a checkout
+```
+
+or, from a built wheel:
+
+```bash
+uv build
+pip install dist/restruct_cv-*.whl
+```
+
+The distribution is `restruct-cv`; the import name and the command it installs are both
+`restruct`. The unqualified name was already taken on PyPI by an unrelated project.
+
+### Model weights
+
+All inference weights are **local-only** and never fetched at run time. Two directories are
+required:
+
+```text
+models/
+  all-MiniLM-L6-v2/
+  distilbert-NER/
+```
+
+They are looked for, in order, in `models/` beside the checkout (when running from one), `models/`
+under the current working directory, and `~/.restruct/models`. Set
+`RESTRUCT_MODELS_DIRECTORY=/path/to/models` to name the directory outright, which is the usual
+answer for an installed copy; when it is set nothing else is consulted. A run with no weights
+exits `20` and names every place it looked.
+
+### Tesseract
+
+Tesseract is a system dependency, needed **only** for scanned pages — a PDF with its own text
+layer and a DOCX never ask for it.
+
+```bash
+brew install tesseract                 # macOS
+apt-get install tesseract-ocr          # Debian/Ubuntu
+```
+
+It is found on `PATH`, or where the installers put it (Program Files on Windows, either Homebrew
+prefix on macOS). PyMuPDF renders an OCR page at 300 DPI and `tesseract` is invoked directly,
+without a shell, using the English language data, LSTM engine mode 1 and page segmentation mode 3.
+Rendered page images are deleted after each document.
+
+## Usage
+
+```bash
+uv run restruct <resume.pdf> -o out.json      # one resume; quiet, writes nothing else
+uv run restruct <resume.docx> -o .            # a directory: writes <resume>.json into it
+uv run restruct <resume.pdf> -o out.json --debug        # + stage 4-5 artifacts in out/
+uv run restruct <resume.pdf> -o out.json --stages 1-3   # + those stages (implies --debug)
+uv run restruct <resume.pdf> -o out.json --reconstruct  # + the result drawn back as a page
+uv run restruct examples/11/resume.json --reconstruct   # draw a result already extracted
+
+uv run restruct                               # batch over resumes-synthetic/
+uv run restruct --truths                      # batch over resumes-truths/ (local, gitignored)
+uv run restruct --unsupported                 # batch over resumes-unsupported/
+```
+
+`--stages` selects **debug artifacts, never whether a pass runs**: every pass feeds the next, so a
+flag that skipped one would quietly produce a different resume.
+
+Exit codes are an API, grouped by decade — `1x` input, `2x` environment, `3x` extraction, `4x`
+output, with `2` left to argparse. `errors.py` names every failure and `cli.py` is the only module
+that maps one to a code, so a caller embedding restruct as a library catches an exception by type
+instead of losing its process.
+
+## Output
+
+`resume.json` is lean and metadata-free: no bounding boxes, fonts, geometry, model names,
+confidences or detection methods. Its first key is `schema_version`, and `resume.schema.json` in
+the repository root is the published contract every fixture is validated against.
+
+Sixteen destinations are always present, always in the same order. A section is `null` when the
+resume has none, `[]` when it exists but yielded no entries; a key is never absent.
+
+```json
+{
+  "schema_version": "1.0",
+  "header_profile": { "name": "…", "emails": ["…"], "phones": ["…"] },
+  "summary": null,
+  "experience": [{ "job_titles": ["…"], "companies": ["…"], "bullets": ["…"] }],
+  "education": [],
+  "…": "…",
+  "others": []
+}
+```
+
+All the evidence behind those values — boxes, fonts, confidences, which method decided what —
+lives in a separate track under `raw/`, and the overlays draw it:
+
+```text
+results/<name>/
+  resume.json           the lean output
+  raw/*.json            the evidence: boxes, fonts, confidences, methods
+  debug/page-N.png      the combined overlay, one per source page
+  debug/pass-*/         passes 1-4, gitignored
+```
+
+The pass-1 dump is named after the reader that produced it — `pymupdf.json` for a PDF, `docx.json`
+for a DOCX, since the two read different things and only one of them has coordinates. The
+single-file form writes it beside the other evidence in `raw/`; the batch writes it to
+`debug/<name>.raw-<reader>.json`.
+
+A model-backed box is drawn more heavily than a deterministic one, so a reader can tell whether a
+box is something the document said or something a model concluded.
+
+### Reading the result back
+
+`--reconstruct` draws `resume.json` back out as a page — `reconstruction.pdf` and one PNG per
+page — so the result can be proof-read by eye.
+
+Drawing a result on its own writes those files flat beside the JSON, named after it
+(`resume-reconstruction.pdf`, `resume-page-1.png`); `-o` names a directory to put them in
+instead, under their plain names.
+
+It answers a different question from the overlays. An overlay draws on top of the document, so it
+shows whether a box landed on the right words; the document keeps making sense regardless of what
+was understood. A reconstruction throws the page away and draws only what was understood, which is
+what makes a bullet filed under education or a date read as a job title visible at a glance.
+
+It is deliberately not a facsimile — imitating the original layout would hide the errors it exists
+to reveal. Absent and empty fields are skipped, so what is on the page is what was extracted, and
+anything the renderer cannot place is drawn in red under UNPLACED rather than dropped. Given a
+`resume.json` as the input it draws that and runs nothing else, loading no models.
+
+### Worked examples
+
+`results/` is regenerated by every run and is not committed. Three resumes are, under
+`examples/` — one per ingestion track, because what each track makes available is different:
+
+| Example | Source | What it shows |
+| --- | --- | --- |
+| [`examples/7.anomaly/`](examples/7.anomaly/) | native PDF, 3 pages | The full evidence track. `debug/page-N.png` draws every section box on the page it came from, model-backed boxes drawn more heavily than deterministic ones. `raw/layout-warnings.json` is written even when empty, so an absent finding is distinguishable from an absent check. |
+| [`examples/9.ocr/`](examples/9.ocr/) | scanned PDF | The same shapes from a page with no text layer. OCR is rebuilt into the line geometry the native path produces, so nothing downstream has an OCR case — which is visible here as an overlay that looks like any other. |
+| [`examples/11/`](examples/11/) | DOCX | No overlays at all. A DOCX has no geometry and must never pretend to, so there is nothing to draw on — the structure comes from what the document states outright: styles, list markers, table cells. |
+
+Each holds the `resume.json`, the `raw/*.json` evidence behind it, and the overlays where there
+are any. They are copied out of `results/` by `tools/refresh_examples.py` rather than curated by
+hand, so they are what the pipeline currently produces and not a snapshot of what it once did.
+Their inputs are in `resumes-synthetic/`, and every one is synthetic.
+
+## Architecture
+
+Five ordered passes over one shared in-memory document.
+
+```text
+ingestion/   physical extraction — native PDF text, per-page OCR fallback, DOCX
+document/    shared types and document-wide statistics
+layout/      row clustering, paragraph/bullet accumulation, unsupported-layout detection
+structure/   heading detection, routing, compound headings, precedence resolver, separators
+parsers/     one module per section shape (header, experience, education, skills, grouped, urls)
+model.py     DistilBERT NER and MiniLM adapters, loaded on first use
+patterns/    deterministic regex evidence, grouped by what it describes
+debug/       artifacts (JSON) and overlays (Pillow), one canvas, one colour registry
+schema.py    the lean, versioned output (contract: resume.schema.json)
+errors.py    the failure taxonomy; only cli.py turns one into an exit code
+pipeline.py  orchestration only — the only module that knows the stage order
+cli.py       argparse and filesystem layout only
+```
+
+Extraction runs in a fixed precedence — deterministic, context-sensitive deterministic, NER,
+MiniLM, geometry, then `other` — and `structure/resolver.py` **enforces** that order rather than
+documenting it: a stage that runs after a weaker one has already claimed a span raises.
+
+v1 targets single-column resumes. Layouts whose reading order cannot be recovered (column gutters,
+vertical text, overlapping boxes, nested tables) are **recorded, never repaired**, in
+`debug/layout-warnings.json`.
+
+## Development
+
+```bash
+uv run pytest                                # full suite
+uv run python -m tests.scorecard             # per-field precision/recall/F1
+uv run restruct && uv run python tools/refresh_examples.py
+git status --short examples/    # clean == byte-identical
+```
+
+See [CONTRIBUTING.md](CONTRIBUTING.md) for the test and scorecard workflow, and `CLAUDE.md` for the
+design decisions behind each module.
+
+Please do not submit **real** resumes or their labels; `resumes-truths/` is gitignored for exactly
+that reason. Fixtures in `resumes-synthetic/` are synthetic and safe to commit.

restruct_cv-0.1.0/README.md

@@ -0,0 +1,198 @@
+# Restruct
+
+Restruct turns messy PDF, DOCX, and scanned resumes into consistent, explainable, versioned JSON.
+It combines OCR, layout geometry, NER, semantic similarity, and deterministic rules so every result
+can be traced back to the thing on the page that produced it.
+
+Built for applicant tracking systems, candidate search, analytics pipelines, talent platforms, and
+other systems that need trustworthy resume data.
+
+---
+
+## Install
+
+```bash
+uv sync                      # from a checkout
+```
+
+or, from a built wheel:
+
+```bash
+uv build
+pip install dist/restruct_cv-*.whl
+```
+
+The distribution is `restruct-cv`; the import name and the command it installs are both
+`restruct`. The unqualified name was already taken on PyPI by an unrelated project.
+
+### Model weights
+
+All inference weights are **local-only** and never fetched at run time. Two directories are
+required:
+
+```text
+models/
+  all-MiniLM-L6-v2/
+  distilbert-NER/
+```
+
+They are looked for, in order, in `models/` beside the checkout (when running from one), `models/`
+under the current working directory, and `~/.restruct/models`. Set
+`RESTRUCT_MODELS_DIRECTORY=/path/to/models` to name the directory outright, which is the usual
+answer for an installed copy; when it is set nothing else is consulted. A run with no weights
+exits `20` and names every place it looked.
+
+### Tesseract
+
+Tesseract is a system dependency, needed **only** for scanned pages — a PDF with its own text
+layer and a DOCX never ask for it.
+
+```bash
+brew install tesseract                 # macOS
+apt-get install tesseract-ocr          # Debian/Ubuntu
+```
+
+It is found on `PATH`, or where the installers put it (Program Files on Windows, either Homebrew
+prefix on macOS). PyMuPDF renders an OCR page at 300 DPI and `tesseract` is invoked directly,
+without a shell, using the English language data, LSTM engine mode 1 and page segmentation mode 3.
+Rendered page images are deleted after each document.
+
+## Usage
+
+```bash
+uv run restruct <resume.pdf> -o out.json      # one resume; quiet, writes nothing else
+uv run restruct <resume.docx> -o .            # a directory: writes <resume>.json into it
+uv run restruct <resume.pdf> -o out.json --debug        # + stage 4-5 artifacts in out/
+uv run restruct <resume.pdf> -o out.json --stages 1-3   # + those stages (implies --debug)
+uv run restruct <resume.pdf> -o out.json --reconstruct  # + the result drawn back as a page
+uv run restruct examples/11/resume.json --reconstruct   # draw a result already extracted
+
+uv run restruct                               # batch over resumes-synthetic/
+uv run restruct --truths                      # batch over resumes-truths/ (local, gitignored)
+uv run restruct --unsupported                 # batch over resumes-unsupported/
+```
+
+`--stages` selects **debug artifacts, never whether a pass runs**: every pass feeds the next, so a
+flag that skipped one would quietly produce a different resume.
+
+Exit codes are an API, grouped by decade — `1x` input, `2x` environment, `3x` extraction, `4x`
+output, with `2` left to argparse. `errors.py` names every failure and `cli.py` is the only module
+that maps one to a code, so a caller embedding restruct as a library catches an exception by type
+instead of losing its process.
+
+## Output
+
+`resume.json` is lean and metadata-free: no bounding boxes, fonts, geometry, model names,
+confidences or detection methods. Its first key is `schema_version`, and `resume.schema.json` in
+the repository root is the published contract every fixture is validated against.
+
+Sixteen destinations are always present, always in the same order. A section is `null` when the
+resume has none, `[]` when it exists but yielded no entries; a key is never absent.
+
+```json
+{
+  "schema_version": "1.0",
+  "header_profile": { "name": "…", "emails": ["…"], "phones": ["…"] },
+  "summary": null,
+  "experience": [{ "job_titles": ["…"], "companies": ["…"], "bullets": ["…"] }],
+  "education": [],
+  "…": "…",
+  "others": []
+}
+```
+
+All the evidence behind those values — boxes, fonts, confidences, which method decided what —
+lives in a separate track under `raw/`, and the overlays draw it:
+
+```text
+results/<name>/
+  resume.json           the lean output
+  raw/*.json            the evidence: boxes, fonts, confidences, methods
+  debug/page-N.png      the combined overlay, one per source page
+  debug/pass-*/         passes 1-4, gitignored
+```
+
+The pass-1 dump is named after the reader that produced it — `pymupdf.json` for a PDF, `docx.json`
+for a DOCX, since the two read different things and only one of them has coordinates. The
+single-file form writes it beside the other evidence in `raw/`; the batch writes it to
+`debug/<name>.raw-<reader>.json`.
+
+A model-backed box is drawn more heavily than a deterministic one, so a reader can tell whether a
+box is something the document said or something a model concluded.
+
+### Reading the result back
+
+`--reconstruct` draws `resume.json` back out as a page — `reconstruction.pdf` and one PNG per
+page — so the result can be proof-read by eye.
+
+Drawing a result on its own writes those files flat beside the JSON, named after it
+(`resume-reconstruction.pdf`, `resume-page-1.png`); `-o` names a directory to put them in
+instead, under their plain names.
+
+It answers a different question from the overlays. An overlay draws on top of the document, so it
+shows whether a box landed on the right words; the document keeps making sense regardless of what
+was understood. A reconstruction throws the page away and draws only what was understood, which is
+what makes a bullet filed under education or a date read as a job title visible at a glance.
+
+It is deliberately not a facsimile — imitating the original layout would hide the errors it exists
+to reveal. Absent and empty fields are skipped, so what is on the page is what was extracted, and
+anything the renderer cannot place is drawn in red under UNPLACED rather than dropped. Given a
+`resume.json` as the input it draws that and runs nothing else, loading no models.
+
+### Worked examples
+
+`results/` is regenerated by every run and is not committed. Three resumes are, under
+`examples/` — one per ingestion track, because what each track makes available is different:
+
+| Example | Source | What it shows |
+| --- | --- | --- |
+| [`examples/7.anomaly/`](examples/7.anomaly/) | native PDF, 3 pages | The full evidence track. `debug/page-N.png` draws every section box on the page it came from, model-backed boxes drawn more heavily than deterministic ones. `raw/layout-warnings.json` is written even when empty, so an absent finding is distinguishable from an absent check. |
+| [`examples/9.ocr/`](examples/9.ocr/) | scanned PDF | The same shapes from a page with no text layer. OCR is rebuilt into the line geometry the native path produces, so nothing downstream has an OCR case — which is visible here as an overlay that looks like any other. |
+| [`examples/11/`](examples/11/) | DOCX | No overlays at all. A DOCX has no geometry and must never pretend to, so there is nothing to draw on — the structure comes from what the document states outright: styles, list markers, table cells. |
+
+Each holds the `resume.json`, the `raw/*.json` evidence behind it, and the overlays where there
+are any. They are copied out of `results/` by `tools/refresh_examples.py` rather than curated by
+hand, so they are what the pipeline currently produces and not a snapshot of what it once did.
+Their inputs are in `resumes-synthetic/`, and every one is synthetic.
+
+## Architecture
+
+Five ordered passes over one shared in-memory document.
+
+```text
+ingestion/   physical extraction — native PDF text, per-page OCR fallback, DOCX
+document/    shared types and document-wide statistics
+layout/      row clustering, paragraph/bullet accumulation, unsupported-layout detection
+structure/   heading detection, routing, compound headings, precedence resolver, separators
+parsers/     one module per section shape (header, experience, education, skills, grouped, urls)
+model.py     DistilBERT NER and MiniLM adapters, loaded on first use
+patterns/    deterministic regex evidence, grouped by what it describes
+debug/       artifacts (JSON) and overlays (Pillow), one canvas, one colour registry
+schema.py    the lean, versioned output (contract: resume.schema.json)
+errors.py    the failure taxonomy; only cli.py turns one into an exit code
+pipeline.py  orchestration only — the only module that knows the stage order
+cli.py       argparse and filesystem layout only
+```
+
+Extraction runs in a fixed precedence — deterministic, context-sensitive deterministic, NER,
+MiniLM, geometry, then `other` — and `structure/resolver.py` **enforces** that order rather than
+documenting it: a stage that runs after a weaker one has already claimed a span raises.
+
+v1 targets single-column resumes. Layouts whose reading order cannot be recovered (column gutters,
+vertical text, overlapping boxes, nested tables) are **recorded, never repaired**, in
+`debug/layout-warnings.json`.
+
+## Development
+
+```bash
+uv run pytest                                # full suite
+uv run python -m tests.scorecard             # per-field precision/recall/F1
+uv run restruct && uv run python tools/refresh_examples.py
+git status --short examples/    # clean == byte-identical
+```
+
+See [CONTRIBUTING.md](CONTRIBUTING.md) for the test and scorecard workflow, and `CLAUDE.md` for the
+design decisions behind each module.
+
+Please do not submit **real** resumes or their labels; `resumes-truths/` is gitignored for exactly
+that reason. Fixtures in `resumes-synthetic/` are synthetic and safe to commit.

restruct_cv-0.1.0/pyproject.toml

@@ -0,0 +1,64 @@
+[project]
+name = "restruct-cv"
+version = "0.1.0"
+description = "Turn messy PDF, DOCX, and scanned resumes into consistent, explainable, versioned JSON"
+readme = "README.md"
+requires-python = ">=3.11"
+keywords = [
+    "resume",
+    "cv",
+    "parser",
+    "ats",
+    "applicant-tracking",
+    "ocr",
+    "pdf",
+    "docx",
+]
+license = "MIT"
+license-files = ["LICENSE"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Developers",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Text Processing",
+    "Topic :: Scientific/Engineering :: Information Analysis",
+]
+dependencies = [
+    "pillow>=11.0",
+    "python-docx>=1.1",
+    "pymupdf>=1.28.2",
+    "sentence-transformers>=6.0.1",
+    "transformers>=4.51.3,<5.14.0",
+]
+
+[[project.authors]]
+name = "Nyein Chan Soe"
+email = "contact@nyeinchansoe.com"
+
+[project.urls]
+Homepage = "https://github.com/Nye1nChanSoe/restruct-cv"
+Repository = "https://github.com/Nye1nChanSoe/restruct-cv"
+Issues = "https://github.com/Nye1nChanSoe/restruct-cv/issues"
+
+[project.scripts]
+restruct = "restruct.cli:main"
+
+[dependency-groups]
+dev = [
+    "pytest>=8.0",
+    "jsonschema>=4.20",
+]
+
+[build-system]
+requires = ["uv_build>=0.12.7,<0.13.0"]
+build-backend = "uv_build"
+
+[tool.uv.build-backend]
+module-name = "restruct"
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = "-q"

restruct_cv-0.1.0/pyproject.toml.orig

@@ -0,0 +1,54 @@
+[project]
+name = "restruct-cv"
+version = "0.1.0"
+description = "Turn messy PDF, DOCX, and scanned resumes into consistent, explainable, versioned JSON"
+readme = "README.md"
+authors = [
+    { name = "Nyein Chan Soe", email = "contact@nyeinchansoe.com" }
+]
+requires-python = ">=3.11"
+keywords = ["resume", "cv", "parser", "ats", "applicant-tracking", "ocr", "pdf", "docx"]
+license = "MIT"
+license-files = ["LICENSE"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Developers",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Text Processing",
+    "Topic :: Scientific/Engineering :: Information Analysis",
+]
+dependencies = [
+    "pillow>=11.0",
+    "python-docx>=1.1",
+    "pymupdf>=1.28.2",
+    "sentence-transformers>=6.0.1",
+    "transformers>=4.51.3,<5.14.0",
+]
+
+[project.urls]
+Homepage = "https://github.com/Nye1nChanSoe/restruct-cv"
+Repository = "https://github.com/Nye1nChanSoe/restruct-cv"
+Issues = "https://github.com/Nye1nChanSoe/restruct-cv/issues"
+
+[dependency-groups]
+dev = [
+    "pytest>=8.0",
+    "jsonschema>=4.20",
+]
+
+[project.scripts]
+restruct = "restruct.cli:main"
+
+[build-system]
+requires = ["uv_build>=0.12.7,<0.13.0"]
+build-backend = "uv_build"
+
+[tool.uv.build-backend]
+module-name = "restruct"
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = "-q"

restruct_cv-0.1.0/src/restruct/__init__.py

@@ -0,0 +1,56 @@
+"""Restruct: explainable resume extraction from PDF and scanned documents.
+
+This module is the public Python API. It re-exports the pipeline entry points
+and the shared document types; everything else lives in the stage packages:
+
+    ingestion/   physical extraction, native and OCR
+    document/    shared types and document representation
+    layout/      rows, paragraphs and bullet reconstruction
+    structure/   headings, key-value pairs and section routing
+    parsers/     one module per section shape
+    models/      NER and embedding adapters
+    patterns/    deterministic evidence
+    debug/       artifacts and overlay rendering
+    schema/      the versioned clean output
+"""
+
+from restruct.document.types import (
+    DetectedHeading,
+    ExtractedLine,
+    HeaderEntityMatch,
+)
+from restruct.schema import build_v1_resume, write_v1_resume
+
+# ``main`` and ``extract_resume`` reach the model libraries, which cost about
+# four seconds to import. Re-exporting them eagerly made every entry into this
+# package pay that -- including `restruct --help` and a run that fails
+# validation before a model is ever consulted. PEP 562 keeps the public names
+# where they were and defers the cost to the first use.
+_LAZY_EXPORTS = {
+    "main": ("restruct.cli", "main"),
+    "extract_resume": ("restruct.pipeline", "extract_resume"),
+}
+
+
+def __getattr__(name: str):
+    target = _LAZY_EXPORTS.get(name)
+    if target is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    import importlib
+
+    module_name, attribute = target
+    return getattr(importlib.import_module(module_name), attribute)
+
+
+def __dir__() -> list[str]:
+    return sorted(__all__)
+
+__all__ = [
+    "DetectedHeading",
+    "ExtractedLine",
+    "HeaderEntityMatch",
+    "build_v1_resume",
+    "extract_resume",
+    "main",
+    "write_v1_resume",
+]