---
license: mit
tags:
  - embeddings
  - skill-retrieval
  - claude-code
---

# superskillret prebuilt index — full-context

Prebuilt embedding index for the [superskillret](https://github.com/ThakiCloud/SUPERSKILLRET) Claude Code plugin.

Unlike the v1 index (which embedded only `name + description`), v2 encodes
the **full skill body** (`name + description + body`) up to
`max_seq_length=32768` tokens. Larger index, much higher recall on
skills whose name/description don't capture every keyword in the body.

- **Version:** 2
- **Corpus:** [`ThakiCloud/SKILLRET`](https://huggingface.co/datasets/ThakiCloud/SKILLRET) (`train+test`)
- **Encoder:** [`ThakiCloud/SkillRet-Embedding-0.6B`](https://huggingface.co/ThakiCloud/SkillRet-Embedding-0.6B)
- **Skills indexed:** 16783
- **Embedding dim:** 1024
- **Encoded text:** `name + description + body` (truncated to 32768 tokens)
- **Normalized:** yes (inner product = cosine similarity)

## Files

| File | Description |
|---|---|
| `skill_embeddings.npy` | FP16 numpy array of shape `(16783, 1024)` |
| `skill_embeddings_int8.npy` | INT8 per-row quantized array of shape `(16783, 1024)` |
| `skill_embeddings_scale.npy` | float32 per-row scale of shape `(16783,)` — reconstruct as `(int8 / 127) * scale` |
| `skill_metadata.jsonl` | one JSON record per row, aligned with embeddings (`name`, `description`, `body`, `source_url`, `namespace`, `repo`, `id`) |
| `VERSION` | integer version tag; bumped when the corpus, encoder, or encoded-text scheme changes |

## Usage

```python
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="ThakiCloud/superskillret-index",
    repo_type="dataset",
    local_dir="cache/",
)
```

Downstream consumers should check `VERSION` against their cached copy before reusing local files.