@cherrystudio/web-search

Portable protocols for the existing Cherry Studio web-search providers. Applications inject networking and explicitly choose provider capabilities; they keep settings, credential storage, key rotation, region detection, retries, fallback, local page extraction, and lifecycle ownership.

import { createWebSearchClient } from '@cherrystudio/web-search'
import { tavily } from '@cherrystudio/web-search/providers/tavily'

const client = createWebSearchClient({
  transport: hostTransport,
  providers: [{ provider: tavily, capabilities: ['searchKeywords'] }]
})

const response = await client.execute({
  providerId: 'tavily',
  capability: 'searchKeywords',
  input: 'Cherry Studio',
  config: { maxResults: 5 },
  credentials: { apiKey: selectedApiKey },
  signal,
  timeoutMs: 30_000
})

hostTransport, selectedApiKey, and signal belong to the application. Each call executes one input against one provider, without implicit retries. Results contain providerId, capability, input, and results with title, url, content, sourceInput, and optional truncated.

Entries

Provider entry exports use their ID as the named export, except exa-mcp, which exports exaMcp. SearXNG additionally exports discoverSearxngEngines; Jina exports getJinaHostCandidates and JINA_MIRROR_HOSTS. Importing the root does not import all protocol modules or content slicing.

SearXNG discovery is explicit and uses the same error, cancellation, and timeout contract as search. Its errors use the searxng/searchKeywords context. Search returns snippets and never automatically downloads result pages. Jina candidate selection is pure; the package does not detect region or retry another endpoint.

apiHost: undefined uses a public default, which for Crawl4AI is the Cloud host https://api.crawl4ai.com. An explicit empty address fails validation. SearXNG has no default deployment. Local fetch and Desktop DuckDuckGo remain host-owned.

Querit accepts API hosts with or without /v1 and numeric-string response codes. Crawl4AI uses the Cloud /scrape protocol for crawl4ai.com and its subdomains, requiring an API key. Other hosts retain the self-hosted /md protocol with an optional token.

Yuanbao rounds maxResults up to the next supported Cnt value (10–50), then applies the requested limit locally. excludeDomains sends up to five distinct domain names through Site; invalid domain entries are omitted. Hosts retain full URL-rule filtering. Yuanbao-only useDefaultResultCount: true omits Cnt for lite-tier compatibility without disabling Site or the local limit. The protocol itself never retries.

Transport and errors

Implement WebSearchTransport.request with the host network stack. Requests contain a complete URL, headers, an already serialized optional body, signal, and optional timeoutMs. Return all HTTP statuses, lowercase response headers, and bodyText; do not parse JSON or encode the body again. Keep URL paths and query encoding intact.

The transport owns timeout enforcement from sending until the body is consumed, forwards cancellation to the actual request, and releases timers/listeners. Omitting timeoutMs retains host defaults. Throw WebSearchTransportError to report network, timeout, cancellation, body-read, or unclassified failures. An abort without an aborted caller signal is not global cancellation.

execute throws WebSearchError. Its kind distinguishes configuration, network, timeout, cancellation, HTTP, provider failure, invalid response, and unknown failure. code, status, providerCode, retryAfter, and requestId support application policy. toJSON() omits causes, raw responses, headers, and credentials. Cancellations retain the AbortError name; timeouts retain TimeoutError. Application-facing messages and retry decisions remain host-owned.

The MCP module supports finite JSON/SSE responses and existing Exa search Text/Highlights and page text formats. Successful result.content: [] or the upstream message No search results found. Please try a different query. means zero search hits. Blank or unrecognized responses are errors, and empty page content is always an error. It does not manage MCP connections, sessions, or streaming subscriptions.

Distribution and development

All public entries ship ESM/CJS and matching declarations, including a react-native export condition. The source has no Node, Electron, React Native, Expo, or application imports. Import resolution checks do not replace consuming-app platform acceptance. Output targets ES2020 and uses standard URL/URLSearchParams and the host cancellation signal; networking stays injected.

Desktop consumes workspace:* as a development dependency through source aliases, so the portable implementation is bundled with the application. Its explicit capability profile filters shared metadata. Mobile must install an exact published version and commit its lockfile. Initial workspace version 0.0.0 is a placeholder; the minor Changeset prepares 0.1.0. During 0.x, breaking API changes require a minor release; compatible fixes use patch releases. Host allowlists prevent new capabilities from being enabled implicitly.

The release workflow explicitly builds and checks this package. test:package checks a built tarball outside the workspace, module formats, declarations, and portable import resolution.

pnpm --filter @cherrystudio/web-search test -- src/__tests__/protocols.test.ts
pnpm --filter @cherrystudio/web-search test:package

These are opt-in development commands. test:package includes compilation and network-backed installation. Follow the current task's authorization rules before running either command.

License: AGPL-3.0-only, preserving the source repository license. Confirm licensing and maintainers before the initial public release.