py-ai-workflows 0.12.0

A toolkit for AI workflows.

The ai_workflows package is a toolkit for supporting AI workflows (i.e., workflows that are pre-scripted and repeatable, but utilize LLMs for various tasks). It’s still in early development, but is ready to support piloting and experimentation.

Installation

Install the latest version with pip:

pip install py-ai-workflows

You’ll also need to install several other dependencies, which you can do by running the initial-setup.ipynb Jupyter notebook — or by installing them manually as follows.

First, download NTLK data for natural language text processing:

# download NLTK data
import nltk
nltk.download('punkt', force=True)

Then install libreoffice for converting Office documents to PDF.

On Linux:

# install LibreOffice for document processing
!apt-get install -y libreoffice

On MacOS:

# install LibreOffice for document processing
brew install libreoffice

On Windows:

# install LibreOffice for document processing
choco install -y libreoffice

Finally, if you’re accessing models via AWS Bedrock, the AWS CLI needs to be installed and configured for AWS access.

Overview

Here are the basics:

1. The llm_utilities module provides a simple interface for interacting with a large language model (LLM). It includes the LLMInterface class that can be used to interact with OpenAI’s models in “JSON mode,” so that you get structured responses parsed to dictionaries for programmatic use.

2. The document_utilities module provides an interface for extracting Markdown-formatted text from various file formats. It includes functions for reading Word, PDF, Excel, CSV, and HTML files, and then converting them into Markdown for use in LLM interactions. Simply create a DocumentInterface object and then call convert_to_markdown() to convert any file to Markdown. If you provide an LLMInterface object, the LLM will be used to help with high-quality conversion.

3. The example-google-colab-document-processing.ipynb notebook provides a simple Google Colab example of how to use the document_utilities module to convert files to Markdown format.

4. The example-google-colab-surveyeval-lite.ipynb notebook provides a more realistic workflow example that uses the document_utilities module to convert a survey file to Markdown format and then to JSON format, and then uses the llm_utilities module to evaluate survey questions using an LLM.

5. The example-surveyeval-lite.ipynb notebook provides the same example, but in a Jupyter notebook that can be run locally.

6. The example-testing.ipynb notebook provides a basic set-up for testing Markdown conversion methods (LLM-assisted vs. not-LLM-assisted).

Typical usage:

from ai_workflows.llm_utilities import LLMInterface
from ai_workflows.document_utilities import DocumentInterface

llm_interface = LLMInterface(openai_api_key=openai_api_key)
doc_interface = DocumentInterface(llm_interface=llm_interface)
markdown = doc_interface.convert_to_markdown(file_path)
dict_list = doc_interface.convert_to_json(
    file_path,
    json_context = "The file contains a survey instrument with questions to be administered to rural Zimbabwean household heads by a trained enumerator.",
    json_job = "Your job is to extract questions and response options from the survey instrument.",
    json_output_spec = "Return correctly-formatted JSON with the following fields: ..."
)

Technical notes

LLMInterface

When you use the LLMInterface class to interact with an LLM, you can provide a formal JSON schema to validate JSON responses. If you only have a human-readable description of the JSON output format, you can convert that to a JSON schema by calling LLMInterface.generate_json_schema(). This method will convert a human-readable description into a formal schema that can be used to validate JSON responses. (The DocumentInterface class will handle this automatically when used to convert documents to JSON. See below for more details.)

By default, the LLMInterface class will retry twice if an LLM response doesn’t parse as JSON or match the schema provided (if any), but you can change this behavior by specifying the json_retries parameter in the constructor.

Currently, the LLMInterface class supports OpenAI models, either directly from OpenAI or via Azure, or the Anthropic models, either directly or via AWS Bedrock.

Markdown conversion

The DocumentInterface.convert_to_markdown() method uses one of several methods to convert files to Markdown:

1. If an LLMInterface is available, PDF files are converted to Markdown with LLM assistance: we split the PDF into pages (splitting double-page spreads as needed), convert each page to an image, and then convert to Markdown using the help of a multimodal LLM. This is the most accurate method, but it’s also the most expensive, running at about $0.015 per page as of October 2024. In the process, we try to keep narrative text that flows across pages together, drop page headers and footers, and describe images, charts, and figures as if to a blind person. We also do our best to convert tables to proper Markdown tables.

2. If an LLMInterface is not available, we use PyMuPDFLLM to convert PDF files to Markdown. This method doesn’t handle images, charts, or figures, and it’s pretty bad at tables, but it does a good job extracting text and a better job adding Markdown formatting than most other libraries. (PyMuPDFLLM also supports a range of other file formats, and we also use it to convert them to Markdown. That includes .xps, .epub, .mobi, .fb2, .cbz, .svg, and .txt formats.)

3. For .xlsx files without charts or images, we use a custom parser to convert worksheets and table ranges to proper Markdown tables. If there are charts or images and we have an LLMInterface available, we use LibreOffice to convert to PDF and, if it’s 10 pages or fewer, we convert from the PDF to Markdown using the LLM assistance method described above. If it’s more than 10 pages, we fall back to the Unstructured method described below.

4. If we have an LLMInterface available, we use LibreOffice to convert .docx, .doc, and .pptx files to PDF and then convert the PDF to Markdown using the LLM assistance method described above. Otherwise, we fall back to the Unstructured method described below.

5. Finally, if we haven’t managed to convert the file using one of the higher-quality methods described above, we use the Unstructured library to parse the file into elements and then add basic Markdown formatting. This method is fast and cheap, but it’s also the least accurate.

JSON conversion

You can convert from Markdown to JSON using the DocumentInterface.markdown_to_json() method, or you can convert files directly to JSON using the DocumentInterface.convert_to_json() method. The latter method will most often convert to Markdown first and then to JSON, but it will convert straight to JSON with a page-by-page approach if:

1. The markdown_first parameter is explicitly provided as False and converting the file to Markdown would naturally use an LLM with a page-by-page approach (see the section above)

2. Or: converting the file to Markdown would naturally use an LLM with a page-by-page approach, the markdown_first parameter is not explicitly provided as True, and the file’s content doesn’t look too large to fit in the LLM context window (<= 50 pages or 25,000 tokens).

The advantage of converting to JSON directly, bypassing the Markdown step, is that you can handle files of arbitrary size. However, the page-by-page approach can work poorly for elements that span pages (since JSON conversion happens page-by-page).

Whether or not you convert to JSON via Markdown, JSON conversion always uses LLM assistance. The parameters you supply are:

1. json_context: a description of the file’s content, to help the LLM understand what it’s looking at

2. json_job: a description of the task you want the LLM to perform (e.g., extracting survey questions)

3. json_output_spec: a description of the output you expect from the LLM

4. json_output_schema: optionally, a formal JSON schema to validate the LLM’s output; by default, this will be automatically generated based on your json_output_spec, but you can specify your own schema or explicitly pass None if you want to disable JSON validation (if JSON validation isn’t disabled, the LLMInterface default is to retry twice if the LLM output doesn’t parse or match the schema, but you can change this behavior by specifying the json_retries parameter in the LLMInterface constructor)

The more detail you provide, the better the LLM will do at the JSON conversion.

If you find that things aren’t working well, try including some few-shot examples in the json_output_spec parameter.

Known issues

LangSmith tracing support is not as good for Anthropic as it is for OpenAI. Specifically: attached images don’t show in the traces and, while token usage does appear in the metadata, the UI doesn’t show token usage nicely.

Roadmap

There’s much that can be improved here. For example:

• Unit testing

• Tracking and reporting LLM costs

• Improving evaluation and comparison methods

• Parallelizing LLM calls for faster processing

• Trying Claude’s direct PDF support

• Adding OCR support for PDF files when an LLM isn’t available

Credits

This toolkit was originally developed by Higher Bar AI, PBC, a public benefit corporation. To contact us, email us at [e-mail]

Full documentation

See the full reference documentation here:

https://ai-workflows.readthedocs.io/

Local development

To develop locally:

1. git clone https://github.com/higherbar-ai/ai-workflows

2. cd ai-workflows

3. python -m venv venv

4. source venv/bin/activate

5. pip install -e .

6. Execute the initial-setup.ipynb Jupyter notebook to install system dependencies.

For convenience, the repo includes .idea project files for PyCharm.

To rebuild the documentation:

1. Update version number in /docs/source/conf.py

2. Update layout or options as needed in /docs/source/index.rst

3. In a terminal window, from the project directory:

   1. cd docs

   2. SPHINX_APIDOC_OPTIONS=members,show-inheritance sphinx-apidoc -o source ../src/ai_workflows --separate --force

   3. make clean html

To rebuild the distribution packages:

1. For the PyPI package:

   1. Update version number (and any build options) in /setup.py

   2. Confirm credentials and settings in ~/.pypirc

   3. Run /setup.py for the bdist_wheel and sdist build types (Tools… Run setup.py task… in PyCharm)

   4. Delete old builds from /dist

   5. In a terminal window:

      1. twine upload dist/* --verbose

2. For GitHub:

   1. Commit everything to GitHub and merge to main branch

   2. Add new release, linking to new tag like v#.#.# in main branch

3. For readthedocs.io:

   1. Go to https://readthedocs.org/projects/ai-workflows/, log in, and click to rebuild from GitHub (only if it doesn’t automatically trigger)