AbdullahElTiby/dsh-vision-bridge ↗★ 0
dsh-vision-bridge-dsh
Vision bridge for text-only models (DeepSeek and any route without image input): Gemini Flash describes attached images at the llm/stream seam and provides the describe_image tool.
AI 分析
核心用途是为不支持图片输入的纯文本模型（如 DeepSeek）桥接视觉能力。适合需要让文本模型理解上传图片的用户，必要条件是需配置 Gemini API 密钥。
インストール
検証済み bundle がないか、互換性チェックに失敗しています。先にリポジトリの説明を読んでください。 README 全文を読む ↗
ドキュメント
README 全文を読む ↗Configuration (row config on the vision-bridge row)
| key | default | meaning |
|---|---|---|
model | gemini-2.5-flash | Gemini vision model id |
apiKeyRef | GEMINI_API_KEY | credential ref (env var / ~/.dsh/.credentials.yaml / .env) |
endpoint | https://generativelanguage.googleapis.com/v1beta | Gemini REST endpoint |
maxOutputTokens | 1024 | description length cap |
temperature | 0.4 | Gemini sampling temperature |
timeoutMs | 30000 | per-call Gemini timeout |
maxImageBytes | 15728640 | largest image sent to Gemini |
admitImages | true | patch resolveModelInfo (image admission) |
tool | true | register describe_image |
systemSection | true | contribute the prompt section |
cacheSize | 256 | description cache size |
Set admitImages: false to keep the stock gates (images are then rejected for
text-only models and the bridge never fires).