brilliant751/dsh-web-search-litellm

README 摘要

@brilliant751/dsh-web-search-litellm English 中文 A LiteLLM /v1/search -backed web search provider for the DeepSeek Harness ( dsh ) web capability seam ( ctx.web ). One search is one HTTP request to LiteLLM's search endpoint — no model turn , unlike the shipped DeepSeek Anthropic-compatible provider. It reuses your LiteLLM gateway and its virtual key, so chat and search go through the same proxy. Install Restart dsh web . The bundle declares dsh.bundle.patch , so dsh plugin add appends it to dsh.profile.bundles , which registers the litellm provider and selects it as the active search backend. How it works - Registers a WebSearchProvider with id litellm into ctx.web . - Calls POST {baseURL}/search with { query, search tool name } and Authorization: Bearer . - Maps LiteLLM's results[] ( title / url / snippet / date ) into the seam's normalized sources[] . - The model-facing web search tool is unchanged ( @deepseek-ai/dsh-tool-web ); results keep the native citation card. Configuration Defaults live in cordis.patch.yml : Key Default Meaning baseURL https://your-litellm.example.com/v1 LiteLLM base; /search is appended. OpenAI-compatible base, not the Anthropic passthrough. apiKeyEnv DEE…

分类