ocrcontext 0.1.3__tar.gz → 0.1.4__tar.gz

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/CHANGELOG.md

@@ -7,7 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
-## [0.1.3] - 2026-06-27
+## [0.1.4] - 2026-06-27
 
 ### Added
 - **GPU acceleration** — `Analyzer(use_gpu=True)` routes PaddleOCR inference to a
@@ -95,7 +95,8 @@ into a standalone, LLM-agnostic library.
 - **Packaging** — optional extras `[paddle]`, `[trocr]`, `[vision]`, `[all]`;
   PEP 561 typed (`py.typed`); examples and a GPU/network-free test suite.
 
-[Unreleased]: https://github.com/bahadirkarsli/ocrcontext/compare/v0.1.3...HEAD
+[Unreleased]: https://github.com/bahadirkarsli/ocrcontext/compare/v0.1.4...HEAD
+[0.1.4]: https://github.com/bahadirkarsli/ocrcontext/compare/v0.1.3...v0.1.4
 [0.1.3]: https://github.com/bahadirkarsli/ocrcontext/compare/v0.1.2...v0.1.3
 [0.1.2]: https://github.com/bahadirkarsli/ocrcontext/compare/v0.1.1...v0.1.2
 [0.1.1]: https://github.com/bahadirkarsli/ocrcontext/compare/v0.1.0...v0.1.1

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ocrcontext
-Version: 0.1.3
+Version: 0.1.4
 Summary: Decoupled, LLM-agnostic document OCR + structured extraction. Vision and LLM parsing in 3 lines of code.
 Project-URL: Homepage, https://github.com/BahadirKarsli/OCRContext
 Project-URL: Repository, https://github.com/BahadirKarsli/OCRContext
@@ -90,8 +90,19 @@ print(result.text)
 
 `ocrcontext` is the extraction core of a production document-analysis platform, lifted out of its FastAPI/Next.js stack into a pure, pip-installable library. It handles OCR engine routing, fidelity-first LLM cleanup, and schema-based structured extraction — and gets out of your way.
 
+## Demo
+
+**Structured invoice extraction from an image:**
+
+<img width="100%" alt="Invoice extraction demo" src="https://github.com/user-attachments/assets/8e77ab83-fff3-4929-9a54-7f4a75acc16f" />
+
+**Digital PDF text extraction:**
+
+<img width="100%" alt="PDF extraction demo" src="https://github.com/user-attachments/assets/84437bd0-9d24-4a2e-8e0c-0014c9e85820" />
+
 ## Contents
 
+- [Demo](#demo)
 - [Install](#install)
 - [CLI](#cli)
 - [Quick start (Python API)](#quick-start-python-api)

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/README.md

@@ -35,8 +35,19 @@ print(result.text)
 
 `ocrcontext` is the extraction core of a production document-analysis platform, lifted out of its FastAPI/Next.js stack into a pure, pip-installable library. It handles OCR engine routing, fidelity-first LLM cleanup, and schema-based structured extraction — and gets out of your way.
 
+## Demo
+
+**Structured invoice extraction from an image:**
+
+<img width="100%" alt="Invoice extraction demo" src="https://github.com/user-attachments/assets/8e77ab83-fff3-4929-9a54-7f4a75acc16f" />
+
+**Digital PDF text extraction:**
+
+<img width="100%" alt="PDF extraction demo" src="https://github.com/user-attachments/assets/84437bd0-9d24-4a2e-8e0c-0014c9e85820" />
+
 ## Contents
 
+- [Demo](#demo)
 - [Install](#install)
 - [CLI](#cli)
 - [Quick start (Python API)](#quick-start-python-api)

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "ocrcontext"
-version = "0.1.3"
+version = "0.1.4"
 description = "Decoupled, LLM-agnostic document OCR + structured extraction. Vision and LLM parsing in 3 lines of code."
 readme = "README.md"
 license = { text = "MIT" }

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/src/ocrcontext/cli.py

@@ -11,6 +11,7 @@ Then run:
 
 from __future__ import annotations
 
+import os
 import sys
 from pathlib import Path
 from typing import Optional
@@ -25,6 +26,7 @@ except ImportError:  # pragma: no cover
 
 from .analyzer import Analyzer
 from .config import AnalyzerConfig
+from .types import OcrResult
 from .schemas import (
     Contract,
     IdCard,
@@ -33,6 +35,61 @@ from .schemas import (
     Receipt,
 )
 
+def _suppress_paddle_noise() -> None:
+    import logging
+    import warnings
+
+    # Set env vars BEFORE any paddle/paddlex import so they see the right paths.
+    # _ensure_ascii_model_cache() in paddle.py does the same but only when the
+    # engine lazy-loads; calling it here guarantees it runs first.
+    from .engines.paddle import _ensure_ascii_model_cache, _ensure_paddle_runtime_flags
+    _ensure_ascii_model_cache()
+    _ensure_paddle_runtime_flags()
+
+    os.environ.setdefault("GLOG_minloglevel", "3")
+
+    # Silence Python-level loggers (no paddlex import — that would defeat the purpose).
+    null = logging.NullHandler()
+    for name in ("ppocr", "paddlex", "paddle", "paddle.utils", "paddle.fluid"):
+        lg = logging.getLogger(name)
+        lg.setLevel(logging.ERROR)
+        lg.handlers = [null]
+        lg.propagate = False
+
+    # Root-level filter catches sub-loggers that bypass the above (e.g. paddlex.utils.*).
+    class _NoiseFilter(logging.Filter):
+        _NOISE = ("Could not find files", "ccache", "oneDNN", "mkldnn")
+        def filter(self, record: logging.LogRecord) -> bool:
+            return not any(t in record.getMessage() for t in self._NOISE)
+
+    logging.getLogger().addFilter(_NoiseFilter())
+
+    warnings.filterwarnings("ignore", category=UserWarning, module="paddle")
+
+
+
+
+def _route_label(result: OcrResult, file_path: Path) -> str:
+    src = result.text_source
+    if src == "pdf_text_layer":
+        return "DIGITAL PDF -> text layer"
+    if src == "ocr":
+        return "SCANNED PDF -> rasterize + PaddleOCR" if file_path.suffix.lower() == ".pdf" else "IMAGE -> PaddleOCR"
+    if src == "vision_handwriting":
+        return "HANDWRITING -> Google Vision"
+    if src == "handwriting_ocr":
+        return "HANDWRITING -> PaddleOCR"
+    return src
+
+
+def _info(msg: str) -> None:
+    typer.echo(f"[i] {msg}", err=True)
+
+
+def _ok(msg: str) -> None:
+    typer.echo(f"[OK] {msg}", err=True)
+
+
 app = typer.Typer(
     name="ocrcontext",
     help="OCR a document and optionally extract structured data.",
@@ -129,12 +186,13 @@ def extract(
 ) -> None:
     """OCR a document and optionally extract structured data."""
 
+    _suppress_paddle_noise()
+
     file_path = Path(file)
     if not file_path.exists():
         typer.echo(f"[ERROR] File not found: {file}", err=True)
         raise typer.Exit(code=1)
 
-    # Validate --schema value early for a clear error message.
     if schema is not None and schema not in _SCHEMAS:
         typer.echo(
             f"[ERROR] Unknown schema '{schema}'. "
@@ -148,36 +206,42 @@ def extract(
         raise typer.Exit(code=1)
 
     refine_flag = _parse_refine(refine)
-
-    # Build LLM only when needed.
     needs_llm = schema is not None or refine_flag is True
     llm = _build_llm(provider, model) if needs_llm else None
 
-    analyzer = Analyzer(
-        llm=llm,
-        config=AnalyzerConfig(lang=lang),
-    )
+    analyzer = Analyzer(llm=llm, config=AnalyzerConfig(lang=lang))
 
     try:
-        if schema is not None:
-            schema_cls = _SCHEMAS[schema]
-            result = analyzer.extract(
-                file_path,
-                schema=schema_cls,
-                handwriting=handwriting,
-                refine=refine_flag or False,
-            )
-            typer.echo(result.model_dump_json(indent=2))
-        else:
-            result = analyzer.analyze(
+        _info(f"file: {file_path.name}")
+        _info("OCR...")
+
+        ocr_result = analyzer.analyze(
                 file_path,
                 handwriting=handwriting,
                 refine=refine_flag,
             )
+
+        conf = f"confidence: {ocr_result.confidence:.0%}" if ocr_result.confidence < 1.0 else "exact"
+        _ok(f"route: {_route_label(ocr_result, file_path)}  ({conf})")
+
+        if ocr_result.refined:
+            _ok("refine: APPLIED")
+
+        if schema is not None:
+            schema_cls = _SCHEMAS[schema]
+            _info(f"extract: {schema} schema...")
+            structured = analyzer.extract_text(
+                ocr_result.text,
+                schema_cls,
+                language=lang,
+            )
+            _ok(f"extract: {schema} [OK]")
+            typer.echo(structured.model_dump_json(indent=2))
+        else:
             if output == "json":
-                typer.echo(result.model_dump_json(indent=2))
+                typer.echo(ocr_result.model_dump_json(indent=2))
             else:
-                typer.echo(result.text)
+                typer.echo(ocr_result.text)
 
     except Exception as exc:  # noqa: BLE001
         typer.echo(f"[ERROR] {exc}", err=True)

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/src/ocrcontext/engines/paddle.py

@@ -104,6 +104,7 @@ class PaddleEngine(OcrEngine):
         import logging
 
         logging.getLogger("ppocr").setLevel(logging.ERROR)
+        logging.getLogger("paddlex").setLevel(logging.ERROR)
         requested = paddle_lang
         ocr, errors = self._try_init(PaddleOCR, paddle_lang, use_gpu=self._use_gpu)
         if ocr is None and paddle_lang != "en":
@@ -129,9 +130,12 @@ class PaddleEngine(OcrEngine):
         # Shared 3.x flags: disable sub-models unneeded for plain OCR.
         # enable_mkldnn is forced False on CPU to avoid PaddlePaddle 3.x PIR bug;
         # on GPU it's irrelevant (MKLDNN is CPU-only) but harmless to keep False.
+        # use_gpu is only injected when True — some PaddleOCR 3.x builds reject it
+        # as an unknown argument even when set to False.
+        gpu_kwargs = {"use_gpu": True} if use_gpu else {}
         base_3x = {
             "lang": lang,
-            "use_gpu": use_gpu,
+            **gpu_kwargs,
             "use_doc_orientation_classify": False,
             "use_doc_unwarping": False,
             "use_textline_orientation": False,
@@ -147,10 +151,10 @@ class PaddleEngine(OcrEngine):
             # 3.x default — version determined by installed package, no pin
             base_3x,
             # Minimal 3.x (for builds that reject the sub-model flags)
-            {"lang": lang, "use_gpu": use_gpu, "enable_mkldnn": False},
-            {"lang": lang, "use_gpu": use_gpu},
+            {"lang": lang, **gpu_kwargs, "enable_mkldnn": False},
+            {"lang": lang, **gpu_kwargs},
             # Legacy 2.x (use_angle_cls; use_doc_* / show_log don't exist in 2.x)
-            {"use_angle_cls": True, "lang": lang, "use_gpu": use_gpu},
+            {"use_angle_cls": True, "lang": lang, **gpu_kwargs},
         ]
         errors: list[str] = []
         for kwargs in profiles:

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/src/ocrcontext/llm/schemas.py

@@ -22,7 +22,7 @@ class LineItem(BaseModel):
             "Default 1 only if neither is available."
         ),
     )
-    unit: Optional[str] = Field(None, description="Unit type (Adet, Kg, Saat, etc.).")
+    unit: Optional[str] = Field(None, description="Unit of measure as written on the document (e.g. pcs, kg, hrs). Null if not present.")
     unit_price: Optional[float] = Field(None, description="Price per unit.")
     tax_rate: Optional[str] = Field(
         None, description="Tax percentage (e.g., 20, 10, 0) or pattern."
@@ -34,9 +34,9 @@ class Invoice(BaseModel):
     supplier_name: Optional[str] = Field(None, description="Name of the vendor/supplier.")
     invoice_date: Optional[str] = Field(None, description="Format YYYY-MM-DD.")
     invoice_number: Optional[str] = Field(None, description="The invoice ID/number.")
-    tax_id: Optional[str] = Field(None, description="Tax ID / VKN / TCKN.")
+    tax_id: Optional[str] = Field(None, description="Tax ID / VAT registration number.")
     tax_rate: Optional[str] = Field(
-        None, description="e.g. 'KDV %20' when KDV is 20%."
+        None, description="Tax/VAT rate as written on the document (e.g. 'VAT 20%', 'GST 10%')."
     )
     currency: Optional[str] = Field(None, description="Currency code (TRY, USD, EUR, etc.).")
     total_amount: Optional[float] = Field(None, description="Final total amount (numeric).")
@@ -63,37 +63,38 @@ class Invoice(BaseModel):
         return self
 
 
-# Verbatim system prompt from app/api/invoices/process/route.ts.
 INVOICE_EXTRACTION_PROMPT = """You are an expert invoice data extraction assistant.
 
 CRITICAL RULES:
-1. **LANGUAGE REPAIR**:
-    - The text may come from OCR and may have missing characters.
-    - If language is 'tr' (Turkish), intelligently fix missing Turkish characters.
+1. **OCR REPAIR**: The text may come from OCR and may have missing or garbled characters.
+   Use context to infer the correct value — do not invent values that are not on the document.
 
 2. **NUMBER PARSING**:
-    - Be extremely careful with comma (,) and dot (.).
-    - In Turkish/European invoices, '1.200,50' means One Thousand Two Hundred and 50 cents.
-    - NEVER confuse a quantity (e.g., 500) with a price (e.g. 5,00).
+    - Be careful with comma (,) and dot (.) as thousand separators vs decimal points.
+    - European format: '1.200,50' = 1200.50. US/UK format: '1,200.50' = 1200.50.
+    - NEVER confuse a quantity (e.g., 2) with a unit price (e.g., 45.00).
 
 3. **CURRENCY DETECTION**:
-    - Look for symbols: ₺, TL, TRY, USD, $, EUR, €.
-    - Prioritize 'TRY' / 'TL' unless explicitly stated otherwise.
+    - Look for symbols or codes on the document: $, USD, €, EUR, £, GBP, ₺, TRY, etc.
+    - Use ONLY what is explicitly stated. Do not default to any currency.
 
-Extract the following fields if it exists:
+4. **UNITS**: Copy the unit exactly as written on the document (pcs, kg, hrs, m², etc.).
+   If no unit is shown, use null — never invent one.
+
+Extract the following fields if present:
 - 'supplier_name': Name of the vendor/supplier.
 - 'invoice_date': Format YYYY-MM-DD.
 - 'invoice_number': The invoice ID/number.
-- 'tax_id': Tax ID / VKN / TCKN.
-- 'tax_rate': It can be like 'KDV' and for example if it is 'KDV' and it is %20, write it as 'KDV %20' in excel.
-- 'currency': Currency code (TRY, USD, EUR, etc.).
+- 'tax_id': Tax ID or VAT registration number.
+- 'tax_rate': Tax/VAT rate as written (e.g. 'VAT 20%', 'GST 10%').
+- 'currency': ISO currency code (USD, EUR, GBP, TRY, etc.).
 - 'total_amount': Final total amount (numeric).
 - 'line_items': An array of items/services. Each item should have:
   - 'description': Product/Service name.
   - 'quantity': Numeric quantity. If missing, calculate it as total / unit_price. Default 1 only if neither is available.
-  - 'unit': Unit type (Adet, Kg, Saat, etc.).
+  - 'unit': Unit of measure exactly as written on the document. Null if not present.
   - 'unit_price': Price per unit.
-  - 'tax_rate': Tax percentage (e.g., 20, 10, 0) or pattern.
+  - 'tax_rate': Tax percentage (e.g., 20, 10, 0) or null.
   - 'total': Total price for this line.
 
 Return ONLY a valid JSON object. If a field is not found, use null."""

{ocrcontext-0.1.3 → ocrcontext-0.1.4}/tests/test_cli.py

@@ -48,6 +48,9 @@ def _patch_analyzer(monkeypatch, ocr_text: str = "hello world", structured: dict
         def extract(self, *args, schema=None, **kwargs):
             return schema(**(structured or {}))
 
+        def extract_text(self, text, schema, **kwargs):
+            return schema(**(structured or {}))
+
     monkeypatch.setattr(cli_mod, "Analyzer", _FakeAnalyzer)
     monkeypatch.setattr(cli_mod, "_build_llm", lambda provider, model: None)
 
@@ -99,7 +102,7 @@ def test_extract_json_output(ascii_tmp, monkeypatch):
     result = runner.invoke(app, ["extract", str(png), "--output", "json"])
     assert result.exit_code == 0
     import json
-    data = json.loads(result.output)
+    data = json.loads(result.output[result.output.index("{"):])
     assert data["text"] == "some text"
     assert data["text_source"] == "ocr"
 
@@ -119,7 +122,7 @@ def test_extract_invoice_schema(ascii_tmp, monkeypatch):
     result = runner.invoke(app, ["extract", str(png), "--schema", "invoice"])
     assert result.exit_code == 0
     import json
-    data = json.loads(result.output)
+    data = json.loads(result.output[result.output.index("{"):])
     assert data["supplier_name"] == "ACME"
     assert data["total_amount"] == 250.0
 
@@ -131,7 +134,7 @@ def test_extract_receipt_schema(ascii_tmp, monkeypatch):
     result = runner.invoke(app, ["extract", str(png), "--schema", "receipt"])
     assert result.exit_code == 0
     import json
-    data = json.loads(result.output)
+    data = json.loads(result.output[result.output.index("{"):])
     assert data["store_name"] == "Migros"