Use gpt-4o-mini to generate doc metadata

[johndmulhausen]

Uses OpenAI's gpt-4o-mini model to generate:

• A description suitable for use in the front-matter of the page (which can also populate <meta name="description" content="{description}">)
• A list of keywords suitable for use in the front-matter of the page (which can also populate <meta name="keywords" content="{keywords}"/>)
• A TL;DR of the page in three bullet points, which can be added to the top of the page

To call, run this from the root of the docs repo, after getting an OpenAI API key at https://platform.openai.com/api-keys

pip install openai
pip install python-frontmatter
OPENAI_API_KEY={project API key} python scripts/generate_metadata.py docs/guides/sweeps/initialize-sweeps.md

Example output:

% OPENAI_API_KEY={redacted}  python scripts/generate_metadata.py docs/guides/sweeps/initialize-sweeps.md

title: Initialize a sweep
description: Initialize a W&B Sweep

NEW DESCRIPTION: Learn how to initialize W&B sweeps using the SDK in Python or the CLI, ensuring you have the appropriate configuration and project settings for optimal results.

NEW KEYWORDS: wandb sweep instructions, sweep configuration, machine learning, Jupyter Notebook, CLI

Summary bullet-points:
 - W&B utilizes a Sweep Controller to manage experimental runs across machines, issuing instructions for new runs to agents on user machines after each sweep completion.
- Users must define a sweep configuration in a YAML file or Python dictionary and ensure that both the W&B Sweep and the W&B Run are in the same project.
- Sweeps can be initialized using either the W&B SDK in Python scripts/Jupyter Notebooks or the W&B CLI, both returning a sweep ID that includes the entity and project names.

[cloudflare-workers-and-pages]

Deploying docodile with    Cloudflare Pages

Latest commit:	7c51095
Status:	✅  Deploy successful!
Preview URL:	https://6cfb1d0b.docodile.pages.dev
Branch Preview URL:	https://ai-generated-metadata.docodile.pages.dev

View logs

[ngrayluna]

Nice. How should we interpret the returned summaries? These two points could be improved, imo

• Users must define a sweep configuration in a YAML file or Python dictionary and ensure that both the W&B Sweep and the W&B Run are in the same project.

  • This part, "...both the W&B Sweep and the W&B Run are in the same project." this is true, but is not really helpful. It'd be cool if ChatGPT used the sentence following that statement " Therefore, the name you provide when you initialize W&B (wandb.init) must match the name of the project you provide when you initialize a W&B Sweep (wandb.sweep)." (See caution blurb on this page).
  • Even better, instead of taking a sentence verbatim, it'd be cool if it could summarize nuances like logging data. The main things to call out w/ the sweep configuration are: defining the method (i.e. search method), metric, and parameter keys, taking note of the key-value pairs w/in the sweep config since those are the values you can create a hyperparameter search w/ later on, sweep search criteria, etc.

• Sweeps can be initialized using either the W&B SDK in Python scripts/Jupyter Notebooks or the W&B CLI, both returning a sweep ID that includes the entity and project names.

  • This is also giving only semi useful/accurate info. The point that ChatGPT is trying to make is (I think) is that there is a different workflow between using the Python SDK within a notebook/Python script vs using a Python script + CLI.

[cloudflare-workers-and-pages]

Deploying docs with    Cloudflare Pages

Latest commit:	9247365
Status:	✅  Deploy successful!
Preview URL:	https://a6724656.docodile.pages.dev
Branch Preview URL:	https://ai-generated-metadata.docodile.pages.dev

View logs