doc-redaction 2.2.4__tar.gz → 2.2.6__tar.gz

{doc_redaction-2.2.4/doc_redaction.egg-info → doc_redaction-2.2.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: doc_redaction
-Version: 2.2.4
+Version: 2.2.6
 Summary: Redact PDF/image-based documents, Word, or CSV/XLSX files using a Gradio-based GUI interface
 Author-email: Sean Pedrick-Case <spedrickcase@lambeth.gov.uk>
 Maintainer-email: Sean Pedrick-Case <spedrickcase@lambeth.gov.uk>
@@ -86,7 +86,7 @@ Follow these instructions to get the document redaction application running on y
 
 ### 1. Package installation
 
-#### Install from source repo (recommended for full features)
+#### Option 1 - Recommended: Install from source repo
 
 Clone the repository and install in editable mode:
 
@@ -96,10 +96,12 @@ cd doc_redaction
 pip install -e .
 ```
 
-##### Full install from source (Paddle and VLM)
+##### Install extras (Paddle or Transformers/Torch VLM)
+
+To install with PaddleOCR:
 
 ```bash
-pip install -e ".[paddle,vlm]"
+pip install -e ".[paddle]"
 ```
 
 Note that the versions of both PaddleOCR and Torch installed by default are the CPU-only versions. If you want to install the equivalent GPU versions, you will need to run the following commands:
@@ -107,6 +109,13 @@ Note that the versions of both PaddleOCR and Torch installed by default are the
 pip install paddlepaddle-gpu==3.2.1 --index-url https://www.paddlepaddle.org.cn/packages/stable/cu129/
 ```
 
+If you want to run VLMs / LLMs with the transformers package:
+
+```bash
+pip install -e ".[vlm]"
+```
+
+
 **Note:** It is difficult to get paddlepaddle gpu working in an environment alongside torch. You may well need to reinstall the cpu version to ensure compatibility, and run paddlepaddle-gpu in a separate environment without torch installed. If you get errors related to .dll files following paddle gpu install, you may need to install the latest c++ redistributables. For Windows, you can find them [here](https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist?view=msvc-170)
 
 ```bash
@@ -114,7 +123,7 @@ pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu129
 pip install torchvision --index-url https://download.pytorch.org/whl/cu129
 ```
 
-#### Install from PyPI (recommended for users and library use)
+#### Option 2 - Install from PyPI
 
 Create a virtual environment (recommended) and install **doc_redaction**.
 
@@ -132,15 +141,21 @@ The package is published on PyPI as **`doc-redaction`** (import name **`doc_reda
 pip install doc_redaction
 ```
 
-Optional extras (same as in `pyproject.toml`):
+Optional extras (same as in `pyproject.toml`). For installing paddleOCR:
+
+```bash
+pip install "doc_redaction[paddle]"
+```
+
+For running VLMs / LLMs with the transformers package:
 
 ```bash
-pip install "doc_redaction[paddle,vlm]"
+pip install "doc_redaction[vlm]"
 ```
 
 For programmatic use (CLI-first API matching Gradio `api_name` routes), see **[Python Package usage (Python)](https://seanpedrick-case.github.io/doc_redaction/src/python_package_usage.html)**. The console script **`cli_redact`** is available after install.
 
-**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running:
+**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running (note that the prerequisites tesseract and poppler will need to be correctly installed following step 2 below):
 
 ```bash
 python -m app
@@ -154,21 +169,21 @@ python -m app
 
 In practice, the **smoothest UI experience** (examples, bundled assets, docs links, predictable relative paths) is still usually via a **repository checkout** or **Docker**, but PyPI install is sufficient to launch the UI as long as you run it from a suitable working folder and have the system dependencies available (or run `python -m doc_redaction.install_deps` first).
 
-#### Docker installation
+#### Option 3 - Docker installation
 
 The doc_redaction Redaction app can be installed by using the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) or Docker compose files ([llama.cpp](https://github.com/ggml-org/llama.cpp), [vLLM](https://docs.vllm.ai/en/stable/)) provided in the repo.
 
-##### Without Llama.cpp / vLLM inference server
-
-If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
-
 ##### With Llama.cpp / vLLM inference server
 
 The project now has Docker and Docker compose files available to pair running the Redaction app with local inference servers powered by [llama.cpp](https://github.com/ggml-org/llama.cpp), or [vLLM](https://docs.vllm.ai/en/stable/). Llama.cpp is more flexible than vLLM for low VRAM systems, as Llama.cpp will offload to cpu/system RAM automatically rather than failing as vLLM tends to do.
 
 For Llama.cpp, you can use the [docker-compose_llama.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_llama.yml) file, and for vLLM, you can use the [docker-compose_vllm.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_vllm.yml) file. To run, Docker / Docker Desktop should be installed, and then you can run the commands suggested in the top of the files to run the servers.
 
-You will need ~40-50GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+You will need ~40 GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+
+##### Without Llama.cpp / vLLM inference server
+
+If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
 
 ### 2. Install prerequisites: Tesseract and Poppler
 

{doc_redaction-2.2.4 → doc_redaction-2.2.6}/README.md

@@ -23,7 +23,7 @@ Follow these instructions to get the document redaction application running on y
 
 ### 1. Package installation
 
-#### Install from source repo (recommended for full features)
+#### Option 1 - Recommended: Install from source repo
 
 Clone the repository and install in editable mode:
 
@@ -33,10 +33,12 @@ cd doc_redaction
 pip install -e .
 ```
 
-##### Full install from source (Paddle and VLM)
+##### Install extras (Paddle or Transformers/Torch VLM)
+
+To install with PaddleOCR:
 
 ```bash
-pip install -e ".[paddle,vlm]"
+pip install -e ".[paddle]"
 ```
 
 Note that the versions of both PaddleOCR and Torch installed by default are the CPU-only versions. If you want to install the equivalent GPU versions, you will need to run the following commands:
@@ -44,6 +46,13 @@ Note that the versions of both PaddleOCR and Torch installed by default are the
 pip install paddlepaddle-gpu==3.2.1 --index-url https://www.paddlepaddle.org.cn/packages/stable/cu129/
 ```
 
+If you want to run VLMs / LLMs with the transformers package:
+
+```bash
+pip install -e ".[vlm]"
+```
+
+
 **Note:** It is difficult to get paddlepaddle gpu working in an environment alongside torch. You may well need to reinstall the cpu version to ensure compatibility, and run paddlepaddle-gpu in a separate environment without torch installed. If you get errors related to .dll files following paddle gpu install, you may need to install the latest c++ redistributables. For Windows, you can find them [here](https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist?view=msvc-170)
 
 ```bash
@@ -51,7 +60,7 @@ pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu129
 pip install torchvision --index-url https://download.pytorch.org/whl/cu129
 ```
 
-#### Install from PyPI (recommended for users and library use)
+#### Option 2 - Install from PyPI
 
 Create a virtual environment (recommended) and install **doc_redaction**.
 
@@ -69,15 +78,21 @@ The package is published on PyPI as **`doc-redaction`** (import name **`doc_reda
 pip install doc_redaction
 ```
 
-Optional extras (same as in `pyproject.toml`):
+Optional extras (same as in `pyproject.toml`). For installing paddleOCR:
+
+```bash
+pip install "doc_redaction[paddle]"
+```
+
+For running VLMs / LLMs with the transformers package:
 
 ```bash
-pip install "doc_redaction[paddle,vlm]"
+pip install "doc_redaction[vlm]"
 ```
 
 For programmatic use (CLI-first API matching Gradio `api_name` routes), see **[Python Package usage (Python)](https://seanpedrick-case.github.io/doc_redaction/src/python_package_usage.html)**. The console script **`cli_redact`** is available after install.
 
-**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running:
+**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running (note that the prerequisites tesseract and poppler will need to be correctly installed following step 2 below):
 
 ```bash
 python -m app
@@ -91,25 +106,25 @@ python -m app
 
 In practice, the **smoothest UI experience** (examples, bundled assets, docs links, predictable relative paths) is still usually via a **repository checkout** or **Docker**, but PyPI install is sufficient to launch the UI as long as you run it from a suitable working folder and have the system dependencies available (or run `python -m doc_redaction.install_deps` first).
 
-#### Docker installation
+#### Option 3 - Docker installation
 
 The doc_redaction Redaction app can be installed by using the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) or Docker compose files ([llama.cpp](https://github.com/ggml-org/llama.cpp), [vLLM](https://docs.vllm.ai/en/stable/)) provided in the repo.
 
-##### Without Llama.cpp / vLLM inference server
-
-If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
-
 ##### With Llama.cpp / vLLM inference server
 
 The project now has Docker and Docker compose files available to pair running the Redaction app with local inference servers powered by [llama.cpp](https://github.com/ggml-org/llama.cpp), or [vLLM](https://docs.vllm.ai/en/stable/). Llama.cpp is more flexible than vLLM for low VRAM systems, as Llama.cpp will offload to cpu/system RAM automatically rather than failing as vLLM tends to do.
 
 For Llama.cpp, you can use the [docker-compose_llama.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_llama.yml) file, and for vLLM, you can use the [docker-compose_vllm.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_vllm.yml) file. To run, Docker / Docker Desktop should be installed, and then you can run the commands suggested in the top of the files to run the servers.
 
-You will need ~40-50GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+You will need ~40 GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+
+##### Without Llama.cpp / vLLM inference server
+
+If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
 
 ### 2. Install prerequisites: Tesseract and Poppler
 
-This application relies on two external tools for OCR (Tesseract) and PDF processing (Poppler). Please install them on your system before proceeding.
+This application relies on two external tools for OCR (Tesseract) and PDF processing (Poppler). Please install them on your system before proceeding. To run the Document Redaction app successfully, these tools need to be installed and either 1. added to PATH, or 2. be in a folder that is directly referenced in the config/app_config.env file with the variables TESSERACT_FOLDER and POPPLER_FOLDER (defined [here](https://github.com/seanpedrick-case/doc_redaction/blob/main/tools/config.py) if you want to see the code). The instructions below will guide you through diffferent ways to install these dependencies.
 
 ---
 

{doc_redaction-2.2.4 → doc_redaction-2.2.6}/README_PYPI.md

@@ -12,7 +12,7 @@ Follow these instructions to get the document redaction application running on y
 
 ### 1. Package installation
 
-#### Install from source repo (recommended for full features)
+#### Option 1 - Recommended: Install from source repo
 
 Clone the repository and install in editable mode:
 
@@ -22,10 +22,12 @@ cd doc_redaction
 pip install -e .
 ```
 
-##### Full install from source (Paddle and VLM)
+##### Install extras (Paddle or Transformers/Torch VLM)
+
+To install with PaddleOCR:
 
 ```bash
-pip install -e ".[paddle,vlm]"
+pip install -e ".[paddle]"
 ```
 
 Note that the versions of both PaddleOCR and Torch installed by default are the CPU-only versions. If you want to install the equivalent GPU versions, you will need to run the following commands:
@@ -33,6 +35,13 @@ Note that the versions of both PaddleOCR and Torch installed by default are the
 pip install paddlepaddle-gpu==3.2.1 --index-url https://www.paddlepaddle.org.cn/packages/stable/cu129/
 ```
 
+If you want to run VLMs / LLMs with the transformers package:
+
+```bash
+pip install -e ".[vlm]"
+```
+
+
 **Note:** It is difficult to get paddlepaddle gpu working in an environment alongside torch. You may well need to reinstall the cpu version to ensure compatibility, and run paddlepaddle-gpu in a separate environment without torch installed. If you get errors related to .dll files following paddle gpu install, you may need to install the latest c++ redistributables. For Windows, you can find them [here](https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist?view=msvc-170)
 
 ```bash
@@ -40,7 +49,7 @@ pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu129
 pip install torchvision --index-url https://download.pytorch.org/whl/cu129
 ```
 
-#### Install from PyPI (recommended for users and library use)
+#### Option 2 - Install from PyPI
 
 Create a virtual environment (recommended) and install **doc_redaction**.
 
@@ -58,15 +67,21 @@ The package is published on PyPI as **`doc-redaction`** (import name **`doc_reda
 pip install doc_redaction
 ```
 
-Optional extras (same as in `pyproject.toml`):
+Optional extras (same as in `pyproject.toml`). For installing paddleOCR:
+
+```bash
+pip install "doc_redaction[paddle]"
+```
+
+For running VLMs / LLMs with the transformers package:
 
 ```bash
-pip install "doc_redaction[paddle,vlm]"
+pip install "doc_redaction[vlm]"
 ```
 
 For programmatic use (CLI-first API matching Gradio `api_name` routes), see **[Python Package usage (Python)](https://seanpedrick-case.github.io/doc_redaction/src/python_package_usage.html)**. The console script **`cli_redact`** is available after install.
 
-**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running:
+**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running (note that the prerequisites tesseract and poppler will need to be correctly installed following step 2 below):
 
 ```bash
 python -m app
@@ -80,21 +95,21 @@ python -m app
 
 In practice, the **smoothest UI experience** (examples, bundled assets, docs links, predictable relative paths) is still usually via a **repository checkout** or **Docker**, but PyPI install is sufficient to launch the UI as long as you run it from a suitable working folder and have the system dependencies available (or run `python -m doc_redaction.install_deps` first).
 
-#### Docker installation
+#### Option 3 - Docker installation
 
 The doc_redaction Redaction app can be installed by using the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) or Docker compose files ([llama.cpp](https://github.com/ggml-org/llama.cpp), [vLLM](https://docs.vllm.ai/en/stable/)) provided in the repo.
 
-##### Without Llama.cpp / vLLM inference server
-
-If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
-
 ##### With Llama.cpp / vLLM inference server
 
 The project now has Docker and Docker compose files available to pair running the Redaction app with local inference servers powered by [llama.cpp](https://github.com/ggml-org/llama.cpp), or [vLLM](https://docs.vllm.ai/en/stable/). Llama.cpp is more flexible than vLLM for low VRAM systems, as Llama.cpp will offload to cpu/system RAM automatically rather than failing as vLLM tends to do.
 
 For Llama.cpp, you can use the [docker-compose_llama.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_llama.yml) file, and for vLLM, you can use the [docker-compose_vllm.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_vllm.yml) file. To run, Docker / Docker Desktop should be installed, and then you can run the commands suggested in the top of the files to run the servers.
 
-You will need ~40-50GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+You will need ~40 GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+
+##### Without Llama.cpp / vLLM inference server
+
+If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
 
 ### 2. Install prerequisites: Tesseract and Poppler
 

{doc_redaction-2.2.4 → doc_redaction-2.2.6}/app.py

@@ -168,6 +168,7 @@ from tools.config import (
     CLOUD_VLM_MODEL_CHOICE,
     COGNITO_AUTH,
     CONFIG_FOLDER,
+    COST_CODE_ACCORDION_OPEN,
     COST_CODES_PATH,
     CSV_ACCESS_LOG_HEADERS,
     CSV_FEEDBACK_LOG_HEADERS,
@@ -315,6 +316,7 @@ from tools.config import (
     SHOW_OCR_GUI_OPTIONS,
     SHOW_PII_IDENTIFICATION_OPTIONS,
     SHOW_QUICKSTART,
+    SHOW_SET_DEFAULT_COST_CODE_BUTTON,
     SHOW_SUMMARISATION,
     SHOW_TRANSFORMERS_LLM_PII_DETECTION_OPTIONS,
     SHOW_WHOLE_DOCUMENT_TEXTRACT_CALL_OPTIONS,
@@ -923,12 +925,13 @@ cost_code_choice_drop = gr.Dropdown(
     value=DEFAULT_COST_CODE,
     label="Choose cost code for analysis",
     choices=[DEFAULT_COST_CODE],
-    allow_custom_value=False,
+    allow_custom_value=True,
     visible=GET_COST_CODES or ENFORCE_COST_CODES,
 )
 set_default_cost_code_button = gr.Button(
     value="Set default cost code",
-    visible=GET_COST_CODES or ENFORCE_COST_CODES,
+    visible=(GET_COST_CODES or ENFORCE_COST_CODES)
+    and SHOW_SET_DEFAULT_COST_CODE_BUTTON,
 )
 
 reset_cost_code_dataframe_button = gr.Button(
@@ -1255,6 +1258,9 @@ with blocks:
     )  # Keeps track of the last page that the annotator was on
     s3_logs_output_textbox = gr.State(value="")
 
+    # Session default cost codes dataframe
+    session_default_cost_codes_df = gr.State(pd.DataFrame())
+
     ## Annotator zoom value
     annotator_zoom_number = gr.Number(
         label="Current annotator zoom level", value=100, precision=0, visible=False
@@ -2334,7 +2340,8 @@ If you are an LLM/agent calling this app programmatically, prefer the **short `g
                                     walkthrough_set_default_cost_code_button = (
                                         gr.Button(
                                             value="Set default cost code",
-                                            visible=show_cost_codes,
+                                            visible=show_cost_codes
+                                            and SHOW_SET_DEFAULT_COST_CODE_BUTTON,
                                         )
                                     )
 
@@ -2786,7 +2793,9 @@ If you are an LLM/agent calling this app programmatically, prefer the **short `g
 
                 if GET_COST_CODES or ENFORCE_COST_CODES:
                     with gr.Accordion(
-                        "Assign task to cost code", open=True, visible=True
+                        "Assign task to cost code",
+                        open=COST_CODE_ACCORDION_OPEN,
+                        visible=True,
                     ):
                         gr.Markdown(
                             "Please ensure that you have approval from your budget holder before using this app for redaction tasks that incur a cost."

{doc_redaction-2.2.4 → doc_redaction-2.2.6}/doc_redaction/install_deps.py

@@ -405,26 +405,26 @@ def _install_windows_poppler(base_dir: Path, force: bool) -> Path:
 
 def _tesseract_latest_installer_url() -> str:
     """
-    Best-effort: scrape the UB Mannheim directory listing and pick the newest
-    64-bit installer filename.
+    Use the GitHub API to get the latest (5.50) tesseract installer URL for Windows.
     """
-    index_url = "https://digi.bib.uni-mannheim.de/tesseract/"
-    req = urllib.request.Request(
-        index_url,
-        headers={"User-Agent": "doc_redaction-install-deps/1.0"},
-    )
+    # Use the GitHub API to get release data
+    api_url = "https://api.github.com/repos/tesseract-ocr/tesseract/releases/tags/5.5.0"
+    req = urllib.request.Request(api_url, headers={"User-Agent": "python-script"})
+
     with urllib.request.urlopen(req) as resp:
-        html = resp.read().decode("utf-8", errors="replace")
+        data = json.loads(resp.read().decode())
 
-    # Example: tesseract-ocr-w64-setup-5.5.0.20241111.exe
-    names = re.findall(r"(tesseract-ocr-w64-setup-[0-9][^\"'>\s]*?\.exe)", html)
-    if not names:
-        raise RuntimeError(
-            "Could not find a tesseract-ocr-w64-setup-*.exe in the listing"
-        )
-    # Sort by string: version/date suffix makes this generally safe for these filenames.
-    newest = sorted(set(names))[-1]
-    return index_url.rstrip("/") + "/" + newest
+    # Search through the 'assets' list for the .exe
+    for asset in data.get("assets", []):
+        name = asset.get("name", "")
+        print("name: ", name)
+        if re.match(r"tesseract-ocr-w64-setup-.*\.exe", name):
+            print("found tesseract download url: ", asset.get("browser_download_url"))
+            return asset.get("browser_download_url")
+        else:
+            print("Not a tesseract installer: ", name)
+
+    raise RuntimeError("Could not find Windows installer in GitHub assets.")
 
 
 def _find_tesseract_bin_from_prefix(prefix: Path) -> Path | None:

{doc_redaction-2.2.4 → doc_redaction-2.2.6/doc_redaction.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: doc_redaction
-Version: 2.2.4
+Version: 2.2.6
 Summary: Redact PDF/image-based documents, Word, or CSV/XLSX files using a Gradio-based GUI interface
 Author-email: Sean Pedrick-Case <spedrickcase@lambeth.gov.uk>
 Maintainer-email: Sean Pedrick-Case <spedrickcase@lambeth.gov.uk>
@@ -86,7 +86,7 @@ Follow these instructions to get the document redaction application running on y
 
 ### 1. Package installation
 
-#### Install from source repo (recommended for full features)
+#### Option 1 - Recommended: Install from source repo
 
 Clone the repository and install in editable mode:
 
@@ -96,10 +96,12 @@ cd doc_redaction
 pip install -e .
 ```
 
-##### Full install from source (Paddle and VLM)
+##### Install extras (Paddle or Transformers/Torch VLM)
+
+To install with PaddleOCR:
 
 ```bash
-pip install -e ".[paddle,vlm]"
+pip install -e ".[paddle]"
 ```
 
 Note that the versions of both PaddleOCR and Torch installed by default are the CPU-only versions. If you want to install the equivalent GPU versions, you will need to run the following commands:
@@ -107,6 +109,13 @@ Note that the versions of both PaddleOCR and Torch installed by default are the
 pip install paddlepaddle-gpu==3.2.1 --index-url https://www.paddlepaddle.org.cn/packages/stable/cu129/
 ```
 
+If you want to run VLMs / LLMs with the transformers package:
+
+```bash
+pip install -e ".[vlm]"
+```
+
+
 **Note:** It is difficult to get paddlepaddle gpu working in an environment alongside torch. You may well need to reinstall the cpu version to ensure compatibility, and run paddlepaddle-gpu in a separate environment without torch installed. If you get errors related to .dll files following paddle gpu install, you may need to install the latest c++ redistributables. For Windows, you can find them [here](https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist?view=msvc-170)
 
 ```bash
@@ -114,7 +123,7 @@ pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu129
 pip install torchvision --index-url https://download.pytorch.org/whl/cu129
 ```
 
-#### Install from PyPI (recommended for users and library use)
+#### Option 2 - Install from PyPI
 
 Create a virtual environment (recommended) and install **doc_redaction**.
 
@@ -132,15 +141,21 @@ The package is published on PyPI as **`doc-redaction`** (import name **`doc_reda
 pip install doc_redaction
 ```
 
-Optional extras (same as in `pyproject.toml`):
+Optional extras (same as in `pyproject.toml`). For installing paddleOCR:
+
+```bash
+pip install "doc_redaction[paddle]"
+```
+
+For running VLMs / LLMs with the transformers package:
 
 ```bash
-pip install "doc_redaction[paddle,vlm]"
+pip install "doc_redaction[vlm]"
 ```
 
 For programmatic use (CLI-first API matching Gradio `api_name` routes), see **[Python Package usage (Python)](https://seanpedrick-case.github.io/doc_redaction/src/python_package_usage.html)**. The console script **`cli_redact`** is available after install.
 
-**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running:
+**Web UI from a PyPI install:** You *can* start the Gradio UI after `pip install doc_redaction` by running (note that the prerequisites tesseract and poppler will need to be correctly installed following step 2 below):
 
 ```bash
 python -m app
@@ -154,21 +169,21 @@ python -m app
 
 In practice, the **smoothest UI experience** (examples, bundled assets, docs links, predictable relative paths) is still usually via a **repository checkout** or **Docker**, but PyPI install is sufficient to launch the UI as long as you run it from a suitable working folder and have the system dependencies available (or run `python -m doc_redaction.install_deps` first).
 
-#### Docker installation
+#### Option 3 - Docker installation
 
 The doc_redaction Redaction app can be installed by using the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) or Docker compose files ([llama.cpp](https://github.com/ggml-org/llama.cpp), [vLLM](https://docs.vllm.ai/en/stable/)) provided in the repo.
 
-##### Without Llama.cpp / vLLM inference server
-
-If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
-
 ##### With Llama.cpp / vLLM inference server
 
 The project now has Docker and Docker compose files available to pair running the Redaction app with local inference servers powered by [llama.cpp](https://github.com/ggml-org/llama.cpp), or [vLLM](https://docs.vllm.ai/en/stable/). Llama.cpp is more flexible than vLLM for low VRAM systems, as Llama.cpp will offload to cpu/system RAM automatically rather than failing as vLLM tends to do.
 
 For Llama.cpp, you can use the [docker-compose_llama.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_llama.yml) file, and for vLLM, you can use the [docker-compose_vllm.yml](https://github.com/seanpedrick-case/doc_redaction/blob/main/docker-compose_vllm.yml) file. To run, Docker / Docker Desktop should be installed, and then you can run the commands suggested in the top of the files to run the servers.
 
-You will need ~40-50GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+You will need ~40 GB of disk space to run everything depending on the model chosen from the compose file. For the vLLM server, you will need 24 GB VRAM. For the Llama.cpp server, 24 GB VRAM is needed to run at full speed, but the n-gpu-layers and n-cpu-moe parameters in the Docker compose file can be adjusted to fit into your system. I would suggest that 8 GB VRAM is needed as a bare minimum for decent inference speed. See the [Unsloth guide](https://unsloth.ai/docs/models/qwen3.5) for more details on working with GGUF files for Qwen 3.5.
+
+##### Without Llama.cpp / vLLM inference server
+
+If you want a working Docker installation without GPU support, you can install from the [Dockerfile](https://github.com/seanpedrick-case/doc_redaction/blob/main/Dockerfile) in the repo. A working example of this, with the CPU version of PaddleOCR, can be found on [Hugging Face](https://huggingface.co/spaces/seanpedrickcase/document_redaction). You can adjust the INSTALL_PADDLEOCR, PADDLE_GPU_ENABLED, INSTALL_VLM, and TORCH_GPU_ENABLED config variables to adjust for PaddleOCR and Transformers packages for local VLM support. Note that GPU-enabled PaddleOCR, and GPU-enabled Transformers/Torch often don't work well together, which is one reason why a Llama.cpp/vLLM inference server Docker installation option is provided below.
 
 ### 2. Install prerequisites: Tesseract and Poppler
 

{doc_redaction-2.2.4 → doc_redaction-2.2.6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "doc_redaction"
-version = "2.2.4"
+version = "2.2.6"
 description = "Redact PDF/image-based documents, Word, or CSV/XLSX files using a Gradio-based GUI interface"
 readme = "README_PYPI.md"
 authors = [

{doc_redaction-2.2.4 → doc_redaction-2.2.6}/tools/config.py

@@ -2569,6 +2569,17 @@ ENFORCE_COST_CODES = convert_string_to_boolean(
 if ENFORCE_COST_CODES:
     GET_COST_CODES = True
 
+COST_CODE_ACCORDION_OPEN = convert_string_to_boolean(
+    get_or_create_env_var("COST_CODE_ACCORDION_OPEN", "True")
+)
+
+SESSION_DEFAULT_COST_CODES_FILENAME = get_or_create_env_var(
+    "SESSION_DEFAULT_COST_CODES_FILENAME", "session_default_cost_codes.csv"
+)
+
+SHOW_SET_DEFAULT_COST_CODE_BUTTON = convert_string_to_boolean(
+    get_or_create_env_var("SHOW_SET_DEFAULT_COST_CODE_BUTTON", "True")
+)
 
 ###
 # WHOLE DOCUMENT API OPTIONS

{doc_redaction-2.2.4 → doc_redaction-2.2.6}/tools/custom_image_analyser_engine.py

@@ -1384,6 +1384,10 @@ def _call_inference_server_vlm_api(
                 # generated so far on each chunk). Printing every chunk reprints completed lines.
                 # Others stream incremental tokens only. Support both.
                 accumulated_response = ""
+                # Qwen3 / vLLM --reasoning-parser: assistant text may stream only in
+                # delta.reasoning_content (or reasoning) while content stays empty. Non-streaming
+                # responses are handled by _extract_choice_message_text; mirror that here.
+                accumulated_reasoning = ""
                 # Track only the current in-progress line (since the last newline) so GUI line
                 # reporting doesn't repeatedly emit the entire response.
                 line_buffer = ""
@@ -1407,7 +1411,30 @@ def _call_inference_server_vlm_api(
 
                             if "choices" in chunk and len(chunk["choices"]) > 0:
                                 delta = chunk["choices"][0].get("delta", {})
-                                token = delta.get("content", "")
+                                token = delta.get("content") or ""
+                                reason_piece = delta.get("reasoning_content")
+                                if not isinstance(reason_piece, str):
+                                    reason_piece = ""
+                                if not reason_piece:
+                                    r_alt = delta.get("reasoning")
+                                    reason_piece = (
+                                        r_alt if isinstance(r_alt, str) else ""
+                                    )
+
+                                if reason_piece:
+                                    if reason_piece == accumulated_reasoning:
+                                        pass
+                                    elif (
+                                        accumulated_reasoning
+                                        and reason_piece.startswith(
+                                            accumulated_reasoning
+                                        )
+                                    ):
+                                        accumulated_reasoning = reason_piece
+                                    else:
+                                        accumulated_reasoning += reason_piece
+                                    output_tokens += 1
+
                                 if not token:
                                     continue
                                 if token == accumulated_response:
@@ -1460,7 +1487,12 @@ def _call_inference_server_vlm_api(
 
                 print()  # newline after stream finishes
 
-                text = accumulated_response
+                _content_stripped = accumulated_response.strip()
+                text = (
+                    accumulated_response
+                    if _content_stripped
+                    else accumulated_reasoning.strip()
+                )
                 stream_elapsed_s = time.perf_counter() - stream_start
 
                 # Try to extract token usage from final chunk if available