Living Whitepaper

An open source experiment tracking how generative AI is used across Aotearoa New Zealand. The site publishes a monthly national "State of AI in New Zealand" snapshot alongside sector-specific living articles. The repository contains two parts:

• src/livingwp – the Python code for an LLM agent that gathers research and writes updates.
• src/website – a Jekyll site that is automatically populated with those updates and served on GitHub Pages.

Contributing

We welcome contributions from the community! There are two main ways you can get involved:

• Open a Pull Request: If you want to make direct edits to the code or documentation, please fork the repository and open a Pull Request with your changes. This includes updates to the agent logic, website, or any other part of the project.

• Open an Issue: If you have suggestions, ideas, or have found a bug, feel free to open an Issue. This is a great way to propose new features, report problems, or discuss improvements.

Editing the Research Prompts

The default industry prompt is defined in src/livingwp/prompts/instructions_research.md. The New Zealand snapshot uses src/livingwp/prompts/instructions_research_nz_snapshot.md.

You can suggest changes to these prompts by either:

• Opening a Pull Request directly with your proposed edits to the prompt file.
• Opening an Issue to discuss or suggest changes to the prompt.

Articles, prompts and models

The settings for each generated article are defined in src/livingwp/config/industries.json. Changing these settings allows you to:

• Add a new industry article. A new page will be created and added to the site the next time the update process runs.
• Configure special non-industry articles such as the nz monthly national snapshot.
• Specify which OpenAI model to use for an article.
• Add a new instructions file to src/livingwp/prompts/ and use it to prompt the research agent for a specific article or industry.

The default runtime now uses the openai-agents Python SDK on the 0.10.x line with gpt-5.4-2026-03-05 as the research model snapshot. You can still override the model with the RESEARCH_MODEL environment variable or per-article config.

We look forward to your contributions!

File search

In addition to web search results, the agent can be prompted to incorporate material from your collated articles, papers and transcripts:

Update the article config

• Create the vector store in platform.openai.com.
• Add a file_store_name key for the article in src/livingwp/config/industries.json.
• Where they're available, you can also provide public URLs for any files in the store by adding filename_urls to the configuration. This will allow the agent to include citation links for any referenced files.

E.g.

   "file_store_name": "finance_files",
   "filename_urls": {        
        "2504.20086v1.pdf": {"title":"Understanding and Mitigating Risks of Generative AI", "url": "https://arxiv.org/pdf/2504.20086v1"},
        "BloombergGPT.pdf": {"title":"BloombergGPT", "url": "https://arxiv.org/pdf/2303.17564"},
        "Large Language Models in Finance.pdf": {"title":"Large Language Models in Finance", "url": "https://arxiv.org/pdf/2311.10723"}
    }

Update the instructions:

• The configuration changes will give the agent access to the required tools but you'll also need to provide specific instructions on how to use them

E.g.

# Gather Information
- ...
- Use the file search tool to perform an extensive review of all of the available curated articles, papers and transcripts.
    
# Output Format
- ...
- For file search results, include a markdown link for every factual claim from a private file. Use the filename_to_link_converter tool to provide the link for each file. Do not include placeholders for file links if the tool doesn't return one. Enclose each file link in parentheses e.g. (markdown_link). Include links for files after any web search links for the same paragraph.
- Ensure all markdown hyperlinks in the final output are correctly formatted
    

Process

flowchart TD
    A[Run Update Task<br/>on Schedule] --> B{Check for<br/>Existing Article}

    B -->|Article Found| C1[Run Deep Research Agent<br/>Topic + Previous Article]
    B -->|No Article| C2[Run Deep Research Agent<br/>Topic Only]

    C1 --> D[Update Article<br/>with Research Results]
    C2 --> D

    D --> E[Commit Changes]

    E -->     F[Open PR<br/>for Review]

    F --> G{Human Review}

    G -->|Needs Edits| H[Human Makes Edits<br/>to PR]
    G -->|Approve| I[Merge PR]

    H --> G

    I --> J[GitHub Action<br/>Auto Deploy to Website]

    J --> K[End]

Loading

Development

Requirements

• uv for Python dependencies
• Ruby and Bundler for running the website locally

Running the Agent

1. Install Python dependencies:

   uv sync

2. Make sure OPENAI_API_KEY is available in your environment before running a real update.

3. Run the agent:

   uv run livingwp

Running the command above iterates over each configured article in src/livingwp/config/industries.json, rewriting the latest page with fresh research and archiving the outgoing version under src/website/whitepaper/content/archive/<slug>/.

That includes both sector articles such as healthcare and the national snapshot article nz.

You can also target specific articles by passing a comma-separated filter. For example:

uv run livingwp nz
uv run livingwp finance,healthcare

Working on the Website

1. Change to the site directory:

   cd src/website

2. Install Ruby gems:

   bundle install

3. Serve the site locally:

   bundle exec jekyll serve

The site will be available at http://localhost:4000 by default.