File size: 6,610 Bytes
ac9f38e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 | ---
license: other
language:
- en
pretty_name: Danbooru Tag Wiki Vector DB
size_categories:
- 1K<n<10K
task_categories:
- sentence-similarity
- feature-extraction
tags:
- danbooru
- anime
- tags
- embeddings
- vector-database
- sqlite
- sqlite-vec
- semantic-search
- harrier-oss
- gemma
---
# Danbooru Tag Wiki Vector DB
A single-file SQLite database of [Danbooru](https://danbooru.donmai.us)
general-category tag wiki pages, with a `sqlite-vec` virtual table holding
640-dim embeddings of each cleaned wiki body. Built to enable natural-language
search over Danbooru's tag vocabulary — give it a phrase like
*"a girl wearing a sailor uniform"* and get back the tags whose wiki
descriptions match.
Source code (fetcher, embedder, query CLI) lives at
[github.com/JackBinary/danbooru-db](https://github.com/JackBinary/danbooru-db).
## At a glance
| | |
|---|---|
| File | `danbooru.db` (single SQLite file, ~36 MB) |
| Tags | 9,322 general-category tags with `post_count >= 1000` and a valid wiki page |
| Embedded | 9,287 tags (a few wiki bodies are empty/stub) |
| Embedding dim | 640 |
| Embedding model | [`mykor/harrier-oss-v1-270m-GGUF`](https://huggingface.co/mykor/harrier-oss-v1-270m-GGUF) (BF16 at index time) — a GGUF of `microsoft/harrier-oss-v1-270m`, a 270M-param Gemma-embedding model with last-token pooling |
| Vector storage | [`sqlite-vec`](https://github.com/asg017/sqlite-vec) `vec0` virtual table |
| Pooling | last-token, L2-normalized |
| Max input | 248 tokens per wiki body (≈1000 chars) — see *Caveats* |
## Schema
Two tables in one SQLite file:
### `tags`
One row per general-category tag.
| column | type | notes |
|---|---|---|
| `rowid` | INTEGER PK | joins to `vec_tags.rowid` |
| `name` | TEXT UNIQUE | e.g. `cat_ears`, `long_hair` |
| `post_count` | INTEGER | Danbooru post count at fetch time |
| `tag_id` | INTEGER | Danbooru tag id |
| `wiki_id` | INTEGER | Danbooru wiki page id |
| `body_raw` | TEXT | Original dtext source from the wiki |
| `body_clean` | TEXT | dtext stripped; `See Also` section extracted; everything from the first `Posts` header onward dropped. **This is what was embedded.** |
| `see_also` | TEXT | JSON array of tag names from the wiki's `See Also` section |
| `other_names` | TEXT | JSON array of alternate names |
| `wiki_updated_at` | TEXT | ISO 8601 |
| `fetched_at` | TEXT | ISO 8601 |
| `embedded_at` | TEXT | ISO 8601, NULL if not embedded |
### `vec_tags`
A `sqlite-vec` virtual table:
```sql
CREATE VIRTUAL TABLE vec_tags USING vec0(embedding float[640]);
```
Keyed by `rowid` matching `tags.rowid`. Vectors are stored as L2-normalized
float32, so cosine similarity equals `1 - distance/2` for the L2 distance
that `sqlite-vec` returns by default.
## Usage
You need the `sqlite-vec` extension loaded into your SQLite connection
(plain SQLite will error on `vec_tags`). In Python:
```python
import sqlite3, sqlite_vec
conn = sqlite3.connect("danbooru.db")
conn.enable_load_extension(True)
sqlite_vec.load(conn)
conn.enable_load_extension(False)
# Plain metadata query — no extension needed for this one:
for name, pc in conn.execute(
"SELECT name, post_count FROM tags ORDER BY post_count DESC LIMIT 5"
):
print(name, pc)
```
Top 5 tags by post count (sanity check):
```
1girl 7884730
solo 6603611
long_hair 5804917
breasts 4638498
looking_at_viewer 4565846
```
### Semantic search
To do retrieval you need to embed a query with the **same model family** as
the index. Harrier expects an instruction prefix for queries (not docs):
```
Instruct: <task>
Query: <text>
```
The companion CLI uses Q8_0 at query time against the BF16 index (cosine
≈ 0.9997 between BF16 and Q8_0 query vectors, so target ranks against the
BF16 corpus are unchanged but Q8_0 is ~5× faster to load and run):
```sh
uv run danbooru-db-query --db danbooru.db "a girl wearing a sailor uniform"
```
The query is L2-normalized and matched with:
```sql
SELECT t.name, t.post_count, v.distance, t.body_clean
FROM vec_tags v
JOIN tags t ON t.rowid = v.rowid
WHERE v.embedding MATCH :query_blob AND k = 10
ORDER BY v.distance;
```
## How it was built
1. **Fetch tags** (`danbooru-db-fetch --phase tags`) — paginated tag list
from Danbooru's API filtered to general category with `post_count >= 1000`.
2. **Fetch wikis** (`danbooru-db-fetch --phase wikis`) — wiki page for each
tag, rate-limited to 1 request/second to be polite. dtext is parsed to
produce `body_clean` (markup stripped, `See Also` extracted to its own
column, content from the first `Posts` header onward dropped).
3. **Embed** (`danbooru-db-embed`) — `body_clean` truncated to 248 tokens
and embedded with the BF16 Harrier-OSS GGUF, L2-normalized, written to
`vec_tags`.
## Caveats
- **248-token truncation.** `llama-cpp-python` hard-caps per-sequence context
at 256 tokens. Wiki bodies are truncated to 248 tokens (≈1000 chars) before
embedding. Tag definitions at the top of each wiki survive; trailing related-tag
lists do not. If you want full-document embeddings, re-embed `body_clean` with
a different runtime.
- **General-category only.** Character/copyright/artist/meta tags are
excluded — this is a vocabulary of *visual content* tags.
- **`post_count >= 1000` floor.** The long tail of rare tags isn't here.
- **Wiki content is a snapshot.** Fetched May 2026. `post_count` and wiki
bodies drift over time; rebuild from the source repo to refresh.
- **Some bodies are empty.** 35 of 9,322 tags have a wiki page but an empty
`body_clean` after cleanup and are not embedded.
## License
The embeddings, schema, and cleaned bodies in this database are derived from
Danbooru's tag wikis, which are user-contributed content on
[danbooru.donmai.us](https://danbooru.donmai.us). Original wiki text remains
the property of its contributors and is subject to Danbooru's terms of use.
The build pipeline (the GitHub repo) is published under its repository
license; this dataset card and the SQLite container are released for research
and personal use. If you redistribute, credit Danbooru and the wiki authors.
## Citation
If this dataset is useful in published work, please cite the embedding model
and the source:
```bibtex
@misc{harrier-oss-v1-270m,
title = {Harrier-OSS-v1-270M},
author = {Microsoft},
url = {https://huggingface.co/microsoft/harrier-oss-v1-270m},
}
@misc{danbooru-tag-wiki-vector-db,
title = {Danbooru Tag Wiki Vector DB},
author = {JackBinary},
url = {https://github.com/JackBinary/danbooru-db},
year = {2026},
}
```
|