[agourakis82]

Summary

• add self-hosted direct OCR mode for pre-cropped inputs via pipeline.enable_layout=false
• add direct_task_type so callers can mark a crop as text, table, or formula without re-running layout detection
• expose the mode through GlmOcr(...), glmocr.parse(...), and CLI --disable-layout --direct-task-type table
• keep a table prompt guardrail to preserve decimal/thousands separators

Why

Issue #241 reports that an isolated table crop can lose decimal commas when it is processed as a fresh document. The issue asks for a way to submit a known region as a table without re-running layout detection on the crop.

This PR adds that explicit path: callers can send the crop directly to OCR as a full-page table region.

CLI:

glmocr parse table_crop.png --mode selfhosted --disable-layout --direct-task-type table

Python:

parser = GlmOcr(mode="selfhosted", enable_layout=False, direct_task_type="table")
result = parser.parse("table_crop.png")

Tests

• /Users/demetriosagourakis/dev/GLM-OCR/.venv/bin/python -m pre_commit run --files glmocr/config.py glmocr/config.yaml glmocr/pipeline/_workers.py glmocr/pipeline/pipeline.py glmocr/api.py glmocr/cli.py glmocr/tests/test_unit.py
• uv run --with pytest -m pytest glmocr/tests/test_unit.py::TestConfig glmocr/tests/test_unit.py::TestPipelineDirectOcr glmocr/tests/test_unit.py::TestGlmOcrParseStream glmocr/tests/test_unit.py::TestGlmOcrConstructor glmocr/tests/test_unit.py::TestPageLoader -q
• uv run --with pytest -m pytest glmocr/tests/test_unit.py -q

[Copilot]

Copilot review overview

🟢 Approval recommended

The change is narrowly scoped, and the test verifies the prompt reaches the request payload.

Review effort: Balanced
Findings: None

What changed in this PR

This PR adds a numeric-fidelity instruction to the packaged self-hosted table-recognition prompt and checks that it reaches outgoing requests.

Changes:

• Ask table recognition to preserve decimal and thousands separators.
• Test that the packaged prompt appears in a table OCR request.

File	Description
glmocr/​tests/​test_unit.py	Tests prompt delivery in the request payload.
glmocr/​config.yaml	Updates the default table prompt.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[agourakis82]

Updated this PR from a prompt-only guardrail into a direct known-crop OCR path.

New behavior:

• pipeline.enable_layout=false skips layout detection in self-hosted mode.
• pipeline.direct_task_type=table sends a pre-cropped table directly to OCR as one full-page table region.
• CLI and Python APIs expose the mode via --disable-layout --direct-task-type table and GlmOcr(..., enable_layout=False, direct_task_type="table").

This now matches the use case from #241: submit an already-known table crop without re-running layout detection on it.