format string | embedding string | dimension int64 | dtype string | count int64 | effective_date string | rules_sha256 string | inputs_sha256 string | vectors_sha256 string | ids list |
|---|---|---|---|---|---|---|---|---|---|
mtg-agentic-cr-vectors-v1 | hf:google/embeddinggemma-2@26711154b641697c443124e316f3d21587f9ff5d:768:qa-titled-v1 | 768 | float32-le | 3,165 | September 25, 2026. | 8d860e451f20f38865b725b42d82feb714c725373dd8f3b32b8652b3eeb070ca | 5cc39d8ada72ff2edac0c2b6eba03b925e4f198f2ea6f95e6832dd76e0eecf07 | bd6bad6741a883c5027b196a0e18ec56fe33dc5e8d00a84b745414a87472f6a3 | [
"CR 100.1",
"CR 100.1a",
"CR 100.1b",
"CR 100.2",
"CR 100.2a",
"CR 100.2b",
"CR 100.2c",
"CR 100.2d",
"CR 100.3",
"CR 100.4",
"CR 100.4a",
"CR 100.4b",
"CR 100.4c",
"CR 100.4d",
"CR 100.5",
"CR 100.6",
"CR 100.6a",
"CR 100.6b",
"CR 100.7",
"CR 101.1",
"CR 101.2",
"CR 101.2a... |
MTG Agentic Comprehensive Rules embeddings
Prebuilt embedding vectors for every numbered rule in the Magic: The Gathering Comprehensive Rules, used by MTG-Agentic-AI for cited rules retrieval. With these vectors, a new install skips embedding the 3,000+ rules locally, which takes minutes on a CPU.
This dataset contains no rule text. It holds rule numbers and float32 vectors only. The program downloads the rules text from Wizards of the Coast, then uses these vectors only after checking that they were built from exactly that text.
Layout
Each index lives in indexes/<rules sha256 prefix>-<model identity sha256 prefix>/:
manifest.json: what the vectors were built from, and the rule number of each row.vectors.f32:count × dimensionlittle-endian float32 values, one row per rule, in manifestidsorder.
| Field | Meaning |
|---|---|
format |
mtg-agentic-cr-vectors-v1 |
embedding |
Model identity: model, pinned revision, dimension and input format |
rules_sha256 |
SHA-256 of the official rules text file |
inputs_sha256 |
SHA-256 of the exact embedded inputs: rule number, title and text, in order |
vectors_sha256 |
SHA-256 of vectors.f32 |
ids |
Rule numbers (CR 702.15b) for each row |
Current index
| Rules effective | Rules | Model | Dimension |
|---|---|---|---|
| September 25, 2026 | 3,165 | google/embeddinggemma-2 @ 26711154b641697c443124e316f3d21587f9ff5d |
768 |
Documents were embedded as title: <section > rule title> | text: <rule text>, queries as
task: question answering | query: <question>, with mean pooling and L2 normalization.
How MTG Agentic uses it
download-rules, draft, review-deck and check-interaction build the local rules index
on first use. Before embedding, they fetch manifest.json for the current rules text and
model. They use the vectors only when the model identity, rules text hash, embedded-inputs
hash, rule numbers and vector checksum all match. Otherwise, or offline, they embed locally.
Choose another dataset with --rules-index OWNER/NAME, or embed locally with
--rules-index none.
To publish an index for new rules:
python -m mtg_agentic export-rules-index --output ./export
hf upload jungter/mtg-agentic-cr-embeddings ./export . --repo-type dataset
License and attribution
Wizards of the Coast, Magic: The Gathering, and their logos are trademarks of Wizards of the Coast LLC. The vectors are derived from the Comprehensive Rules, © Wizards of the Coast LLC. This is unofficial Fan Content permitted under the Fan Content Policy. It is not approved or endorsed by Wizards.
The vectors were produced with google/embeddinggemma-2, released under Apache 2.0.
- Downloads last month
- 49