dsh-plugin-nuaa-search
A zero-key web search provider for DeepSeek Harness's built-in web_search tool.
It registers a search provider (local-search) into the harness's ctx.web seam and
answers queries by scraping real search engines — no official DeepSeek API key, no
third-party search API key, no settings card.
中文说明见 README.zh.md。
Third-party community plugin, not affiliated with Nanjing University of Aeronautics and Astronautics. The nuaa in the name only helps students find it in the market; the plugin itself is network-agnostic.
Why
@deepseek-ai/dsh-web-search-deepseek sends every search to DeepSeek's official
Anthropic-compatible endpoint, which needs an official API key. On a network or plan where
that key is unavailable (for example a mainland-China campus gateway that only serves an
OpenAI-compatible endpoint), web_search fails. This plugin replaces the provider instead
of the endpoint: it implements the harness's own WebSearchProvider interface and fetches
results itself.
Install
This package is not published to npm (npm signup was unavailable for the author), so
install it from the market or from a prebuilt release tarball — do not run
dsh plugin add <package-name>, there is no registry entry behind that name.
From the market (easiest): open DSH → Settings → Plugins, search nuaa, and install.
Listing: https://awesome-dsh-plugin.com/p/Oxygen640/dsh-plugin-nuaa-search/
From the prebuilt tarball:
dsh plugin --profile desktop add "https://github.com/Oxygen640/dsh-plugin-nuaa-search/releases/latest/download/dsh-plugin-nuaa-search.tgz"
From source:
dsh plugin --profile desktop add github:Oxygen640/dsh-plugin-nuaa-search
Restart the harness. The bundle patch points web.searchProvider at local-search and
disables the official web-search-deepseek row (re-enable it in your profile's
cordis.patch.yml if you want it back).
Backends
auto (default) = bing → so360 → searxng: the first backend that returns results
wins, so a blocked or broken engine costs one timeout, not the whole search.
Query variants. Chinese long-tail queries break Bing's ranking: 南京航空航天大学图书馆
returns the official library pages, while 南京航空航天大学图书馆 预约座位 returns generic
"南京市" pages instead. So each search also runs a second variant with trailing intent words
(开放时间, 预约座位, 几点, 怎么…, 地址, …) stripped, and keeps whichever result set
is more relevant to the entity phrase (longest-common-substring coverage; ties keep the
original query so ordinary searches are unaffected). Measured on the failing case: 1/8 → 7/8
results from nuaa.edu.cn. Turn it off with stripIntent: false.
| Backend | Status | Notes |
|---|
bing | default, ~270ms | Bing China HTML: direct URLs, snippets, publish dates |
so360 | fallback | 360 Search HTML: direct URLs |
searxng | needs your instance | JSON API; the instance must enable format=json |
baidu | often blocked | returns [] when the anti-bot page appears, never the challenge page |
Config
Passed as the row config in cordis.patch.yml, or via environment variables as fallback.
| Key | Default | Meaning |
|---|
backend | auto | one backend or a chain ("bing,so360") |
maxResults | 8 | cap; the caller's request.maxResults is also honored |
deadlineMs | 20000 | internal deadline — returns what it has instead of hanging |
backendTimeoutMs | 8000 | per-backend timeout |
searxngUrl | — | instance base URL |
userAgent | Chrome 126 UA | user agent used for scraping |
log | true | one summary line per search in the host log |
stripIntent | true | also search the entity phrase with trailing intent words stripped |
Environment fallbacks: SEARCH_BACKEND, SEARCH_MAX_RESULTS, SEARCH_DEADLINE_MS,
SEARXNG_URL, SEARCH_USER_AGENT.
Test
node --test test/provider.test.mjs
17 tests: backend-chain and config logic, the provider contract, query-variant relevance,
live network searches, maxResults, deadline, cancellation (including an already-aborted
signal), and query length normalization.
Releasing a new version
No registry is involved, so a release is: bump version in package.json, then build and
attach the tarball under the same asset name so the latest/download/ URL keeps working.
npm pack --pack-destination release
ren release\dsh-plugin-nuaa-search-0.1.2.tgz dsh-plugin-nuaa-search.tgz
Then publish a GitHub release and attach dsh-plugin-nuaa-search.tgz.
Limitations
- HTML scraping: a Bing markup change needs a parser update (switch
backend to so360 meanwhile).
- Bing ranks multi-token Chinese queries badly. It handles a clean entity name
(
南航图书馆, 南京航空航天大学图书馆) well, but degrades toward generic city pages when
extra qualifiers pile up — 南京航空航天大学将军路校区图书馆 already returns "南京市".
The query-variant fix covers the common "entity + intent word" shape, not this one.
site: queries are not supported by the HTML endpoint and are passed through as plain text.- Relevance is the search engine's; results are passed through, not rewritten or summarized.
publishedAt is only set when a date is readable at the start of the snippet — never invented.- One query per call, no pagination.
- Queries are truncated to 200 characters, which is where search engines stop honoring them anyway.
License
MIT
