Usage

In any session, open the model selector and pick:
- DeepSeek (视觉桥) → DeepSeek-V4-Flash (视觉桥) — recommended: coding + native image UX, vision supplied by the bridge.
- Official DeepSeek route — paste falls back to path text; bridge tools remain available.
Paste or drop an image. A thumbnail appears (twin) or a path is inserted (official).
Ask normally. DeepSeek answers from the bridge analysis; use query_region / extract_text for coordinate-level detail.

To make the twin the default for every new session:


## Configuration

| Variable | Default | Purpose |
| --- | --- | --- |
| `OLLAMA_BASE_URL` | `http://127.0.0.1:11434/v1` | Vision engine endpoint (OpenAI-compatible) |
| `OLLAMA_VISION_MODEL` | `qwen2.5vl:7b-q3_K_M` | Vision model used by the bridge tools |
| `OLLAMA_API_KEY` | `ollama` | Compatible key for the engine |
| `VISION_MAX_TOKENS` / `VISION_TEMPERATURE` / `VISION_MAX_RETRIES` / `VISION_CONCURRENCY` | `8192` / `0.1` / `1` / `1` | Inference parameters |
| `DEEPSEEK_API_KEY` | credentials service | Twin-route key — same source as the official route |
| `DEEPSEEK_BASE_URL` | public API | Twin-route endpoint — same resolution as the official route |

Tool rows are disabled by setting `pasteToPath: false` on the bundle row if the paste router is
undesired on a specific deployment.

@yulee-314/dsh-vision-bridge

AI 分析

安裝

說明文件

Usage