Roadmap info!

[benjaminapetersen]

Browsing and noticed that 2014 and 2024 have differences, which makes sense given 2024 is new. I believe a ROADMAP could be handy for folks to help out & know what is already being worked on. I see:

// 2014 resources
[
  "ability_scores",
  "alignments",
  "backgrounds",
  "classes",
  "conditions",
  "damage_types",
  "equipment",
  "equipment_categories",
  "feats",
  "features",
  "languages",
  "levels",
  "magic_items",
  "magic_schools",
  "monsters",
  "proficiencies",
  "races",
  "rule_sections",
  "rules",
  "skills",
  "spells",
  "subclasses",
  "subraces",
  "traits",
  "weapon_properties"
]
// 2024 resources
[
  "ability_scores",
  "alignments",
  "backgrounds",
  "conditions",
  "damage_types",
  "equipment",
  "equipment_categories",
  "feats",
  "languages",
  "magic_schools",
  "proficiencies",
  "skills",
  "weapon_mastery_properties",
  "weapon_properties"
]

I bet Monsters will be quite the undertaking, given the 2024 changes.

[bagelbits]

Here is a rough roadmap of what needs to be done. I just haven't had any time to do it:

[benjaminapetersen]

Very nice.
Is there a canonical source material you are using to convert to JSON? Markdown files or something, if one would like to contribute to this effort for the 2024 API material?

[bagelbits]

@benjaminapetersen Unfortunately, it's a PDF.

[mixtli]

If I could make a suggestion... if we could get schemas for the unimplemented types so we know how you want to structure it, community could probably add items.

[benjaminapetersen]

I have another offer~
I have been playing with the 2014 data quite a bit, alongside the 2024 data, working on a local app in golang and a frontend in typescript. I've made Schemas in both golang and typescript to ensure frontend and backend are strongly typed and match the 2014 data.

I have access to Claude, I'm considering carefully leading the model through the 2024 data instructing it to suck the content from the PDF and put it into JSON that matches, roughly, the 2014 when possible, and deviates into a consistent structure (defined) where the 2024 does not match. For example, Monsters have changed like this:

• Legacy 2014 Adult Blue Dragon: https://www.dndbeyond.com/monsters/16765-adult-blue-dragon
• 2024 Adult Blue Dragon: https://www.dndbeyond.com/monsters/5194870-adult-blue-dragon

I think this is a fantastic use of an LLM's strengths that, critically, isn't actually taking a human's artistry. It's just data processing. Would this be a welcome contribution (or set of smaller contributions)?

For extra effort, I've made a set of YAML files filled with restricted terms (like Beholder, Mind Flayer, DnD, etc) that cannot be included in th final output, and have a second instance scanning the JSON docs to ensure none of these term violations appear in the output due to LLM halucinations.

FWIW, I'm likely to do this for my own usage regardless, but am happy to open a PR here if it would be a welcome contribution & save community effort years of tedious data structuring.

@bagelbits thoughts?

[ogregoire]

Technically, the terms you mention are present in the SRD 2014, and are open, the associated stat blocks aren't, however.

If you use Claude to extract the data, you're never sure there are no hallucinations (incorrect data, prohibited name, etc.). If you use Claude to create an extractor, and review thoroughly the extractor, you actually created a data processor. I'm always of the idea that creating a processor is better because it's produces reproducible results. Ask twice the same prompt in different clean sessions of the same LLM, and you'll get different results.

Also, extracting from a PDF is hard: lots of issues, lots of noise, etc. Some people have already converted the SRD into markdown. Maybe built upon that for better accuracy?

https://github.com/springbov/dndsrd5.2_markdown
https://github.com/downfallx/dnd-5e-srd-markdown

Anyways, sorry to have hijacked this. I got a mail notification for this thread, but I have no clue why.

[benjaminapetersen]

Not a hijack, very relevant. Worth the conversation!

[benjaminapetersen]

So, given there are 2 repos with Markdown already created... thoughts on utilizing that to finish out the 2024 dataset here?

[bagelbits]

Markdown should definitely be used in descriptions where possible.

As an update, our current roadmap is as follows.

flowchart TD
    level["level: Class,Subclass,Feature,Spell"]
    feature["feature: Level,Subclass"]
    spell["spell: AbilityScore,DamageType,MagicSchool,Class,Subclass"]
    monster["monster: AbilityScore,Condition,Proficiency,Spell,Monster"]
    ruleSection
    rule["rule: RuleSection"]
    level --> feature
    spell --> monster
    ruleSection --> rule

Loading

There is a draft PR for feature, but it may be merged in soon. monsters has a few entries added, but could use more.

[G-Ambatte]

Update:

• Features and Levels are now complete
• Monsters and Spells are not yet started/incomplete; from memory, these are the two biggest data sets
• Rules, Rule Sections, and Traps are also not yet started, but are much smaller sets

I believe that Monsters is both the most complex and the largest data set left to be tackled.