dsh-web-search-searxng
A SearXNG-backed WebSearchProvider for the DeepSeek Harness web
capability seam (ctx.web).
It registers a provider with the id searxng and adds no tool of its own: web_search is already
registered by @deepseek-ai/dsh-tool-web, and this package only gives it a backend that needs no
third-party API key.
Why this exists
The default provider (deepseek-official) authenticates in one of two ways:
- a DeepSeek account sign-in, used only when the session's model request is served by the
deepseek-accountprovider route; - a
DEEPSEEK_API_KEY.
When models are served through a self-hosted OpenAI-compatible gateway, (1) never applies and (2)
may not be configured, so every search fails with WEB_PROVIDER_CREDENTIAL_MISSING. Pointing the
same web_search tool at a local SearXNG instance removes that dependency entirely.
Requirements
A running SearXNG instance with the JSON output format enabled — it is off by default:
search: formats: - html - jsonNode's built-in
fetch(Node.js >= 18). No runtime dependencies.
Install
Inside the profile you want the provider in:
dsh plugin --profile web add link:/path/to/dsh-web-search-searxng
Then add two entries to that profile's cordis.patch.yml:
- insert:
- id: web-search-searxng
name: dsh-web-search-searxng
config:
baseURL: http://127.0.0.1:8888
timeoutMs: 20000
# Select it. Profile patches replace a row's config rather than deep-merging,
# so restate every key the row needs.
- id: web
name: '@deepseek-ai/dsh-web'
config:
searchProvider: searxng
fetchProvider: http
Configuration
| Key | Default | Description |
|---|---|---|
baseURL |
http://127.0.0.1:8888 |
SearXNG origin; a trailing slash is tolerated |
language |
— | Forwarded as SearXNG's language (e.g. zh-CN); omit for the default all |
safesearch |
0 |
0 off, 1 moderate, 2 strict |
timeoutMs |
20000 |
Per-request budget |
Failure behaviour
| Situation | Result |
|---|---|
| Instance unreachable | Throws with the origin and a docker ps / colima start hint |
| Non-2xx response | Throws with the status code and origin |
200 with zero results and failing engines |
Throws, including unresponsive_engines |
200 with zero results and no engine errors |
Returns an empty sources array — a genuine miss |
The third row is deliberate. SearXNG reports "no matching results" and "every engine failed" with
the same 200/empty payload, so a passthrough makes a broken deployment look like a quiet miss.
Engine selection and rate limits
SearXNG is a metasearch front end: most engines ship disabled, and which ones answer depends on your network. Two engines are not enough. When Brave or Google throttles you, SearXNG suspends that engine for about 180 seconds (CAPTCHA blocks last far longer), and with no surviving engine the search returns nothing at all.
Keep at least two independent upstreams enabled, for example:
engines:
- name: yandex
disabled: false
- name: 360search
disabled: false
Measure instead of guessing — this lists the engines that failed on one query:
curl -sG http://127.0.0.1:8888/search \
--data-urlencode 'q=test' --data 'format=json' | jq .unresponsive_engines
The provider reports throttling as such rather than as a deployment fault: when every result is empty and the failing engines mention rate limits, the error says SearXNG retries the engine after a ~180s cooldown while the upstream quota can take considerably longer — so retrying immediately tends to fail again, and the useful responses are waiting or enabling a fallback engine.
Example instance
examples/searxng/ contains a minimal docker-compose.yml and settings.yml (JSON enabled,
limiter off). Copy them, set a secret_key, adjust the outbound proxy if your network needs one,
then docker compose up -d.
Uninstall
Remove the insert entry and point web.searchProvider back at the provider you want. The package
can stay installed; nothing else references it.
No comments yet. Be the first to write one.