dsh-llm-vision-bridge

DeepSeek vision bridge for the dsh web GUI: route image attachments to a vision model (pi-ai / llama.cpp Qwen3-VL) and continue the conversation with the text description on a text-only LLM (DeepSeek)

AI 分析

核心用途是让纯文本模型（如 DeepSeek 官方接口）间接具备“看图”能力。适合需要对上传图片进行分析，但又想继续使用强大纯文本模型进行深度对话的用户。

パッケージ

dsh-llm-vision-bridge

バージョン

0.1.2

ライセンス

MIT

最終更新

2026/08/14

Configuration (Settings → Models → llm-vision-bridge)

Field	Default	Description
`enabled`	`true`	Master switch; when off the bridge provider degrades to pure passthrough
`autoRoute`	`false`	Additionally rewrite `deepseek-official` agent requests to the bridge provider (cannot bypass image admission; fallback only)
`fallbackProvider`	`deepseek-official`	The text-only provider that actually generates the reply
`visionProvider`	`llama`	Vision provider route (pi-ai)
`visionModel`	`/models/qwen3-vl-4b-thinking/Qwen3-VL-4B-Thinking-Q4_K_M.gguf`	Vision model id
`visionPrompt`	(built-in Chinese prompt)	System prompt for the vision model
`visionMaxTokens`	`2048`	Vision output cap (keep ≥1024; thinking consumes tokens)
`visionRetries`	`3`	Max retries for retryable errors (503/429/timeout)
`visionRetryDelayMs`	`30000`	Retry delay
`onVisionFailure`	`placeholder`	Final failure policy: `placeholder` = insert a failure note and continue; `error` = fail the turn

dsh-llm-vision-bridge

AI 分析

インストール

ドキュメント

Configuration (Settings → Models → llm-vision-bridge)