@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
dsh plugin --profile web add @brilliant751/dsh-web-search-litellm
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.
LiteLLM prerequisite
LiteLLM routes search through a named search tool. The name is local to your
gateway—it is not the provider type. The bundled default expects a tool named
google. For example, a LiteLLM config.yaml can map that name to Google PSE:
search_tools:
- search_tool_name: google
litellm_params:
search_provider: google_pse
api_key: os.environ/GOOGLE_PSE_API_KEY
Google PSE also expects GOOGLE_PSE_ENGINE_ID in the LiteLLM server environment.
Other supported search_provider values can use the same local tool name.
Before configuring DSH, verify that the gateway key can see the tool:
curl https://your-litellm.example.com/v1/search/tools \
-H 'Authorization: Bearer YOUR_LITELLM_VIRTUAL_KEY'
The response's data[].search_tool_name must contain the name configured below.
How it works
- Registers a
WebSearchProvider with id litellm into ctx.web.
- Calls
POST {baseURL}/search with { query, search_tool_name } and Authorization: Bearer <key>.
- 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 | DEEPSEEK_API_KEY | Credential reference, resolved per search through ctx.credentials (the Models page writes it), then the DSH launch environment. |
searchToolName | google | The LiteLLM search tool to invoke (see GET {baseURL}/search/tools). |
The package exports a Cordis Config schema, so compatible DSH configuration
surfaces can render these fields and keep a literal apiKey redacted. The
bundle patch remains the portable source of initial values.
Override any value from your profile's own cordis.patch.yml or a --patch overlay. A patch replaces a targeted row's whole config (it does not merge), so restate the complete config when overriding:
- insert:
- id: web-search-litellm
name: '@brilliant751/dsh-web-search-litellm'
config:
baseURL: https://your-litellm.example.com/v1
apiKeyEnv: DEEPSEEK_API_KEY
searchToolName: google
To switch back to the shipped DeepSeek provider, override web.searchProvider back to deepseek-official (or uninstall this bundle).
Credential resolution
- A literal
apiKey in the config (not recommended — prefer the credential store).
ctx.credentials resolving apiKeyEnv (the Web UI's Models/Web-search card writes $DSH_HOME/.credentials.yaml).- The DSH launch-environment snapshot: inherited process variables, then the
project
.env, then the Harness home .env.
Compatibility
Version 0.1.1 supports the DSH web seam prerelease lines from
0.1.0-rc.6 through 0.1.2-rc.1 and requires Node.js 20 or newer.
Troubleshooting
- Configured provider is unavailable: check that
baseURL is an absolute
URL, searchToolName is non-empty, and apiKeyEnv resolves in the credential
store or launch environment.
- HTTP 401/403: the virtual key must be allowed to call the selected LiteLLM
search tool; this permission is separate from chat model access.
- Unknown search tool/model: call
GET {baseURL}/search/tools and copy an
exact search_tool_name into the plugin config.
- Invalid search response: confirm the configured route is LiteLLM's Search
API and that the gateway returns a
results array.
Security note
Installing this plugin runs its code with your permissions. It sends your LiteLLM key to the configured baseURL only. Point baseURL only at a gateway you trust.
License
MIT
