dsh-deepseek-exa-hybrid-search

dsh-deepseek-exa-hybrid-search

A DeepSeek Harness (dsh) search provider that queries DeepSeek Search API and Exa at the same time and gives the agent one merged, de-duplicated result list instead of two overlapping ones.

一个把 DeepSeek 与 Exa 的搜索结果并发查询、去重合并成单份结果的 DSH 插件。它替换内建 web_search 的后端，不新增工具。

• One web_search call → both engines in parallel.
• Cross-source hits first, then reciprocal-rank (RRF) ordering.
• One engine down/slow → the other leg still answers (no failed search).
• No extra tools, no client bundle: the only footprint is the search provider.

Why a provider (and not MCP)

ctx.web accepts exactly one provider id per capability and has no multi-source merge. A hosted MCP server can add another tool, but it cannot rewrite the built-in web_search or return one combined list. So the merge lives in a provider that calls both HTTP APIs directly.

Install

dsh plugin --profile web add dsh-deepseek-exa-hybrid-search

Restart the engine after installing. The plugin registers hybrid2 and takes over the DeepSeek search slot (hybrid2 / deepseek-official / deepseek-search-api / unset). Any other explicit provider (for example ddg) is left alone.

Configure

Two keys, set in Settings → Plugins → dsh-deepseek-exa-hybrid-search → Settings (native form), or directly in the profile patch:

- id: web-search-hybrid
  config:
    deepseekApiKey: "sk-..."
    exaApiKey: "..."

Both are plain text in the profile patch by design. That means dsh --dump-config, the settings form, and config backups show them — do not share your patch or a config dump. Precedence:

1. deepseekApiKey / exaApiKey (config)
2. the credentials service (DEEPSEEK_API_KEY / EXA_API_KEY)
3. the process environment

Either leg alone is enough; with only one key the search degrades to that leg.

Behavior

• DeepSeek: POST https://api.deepseek.com/v1/tools/search (structured results, no model turn).
• Exa: POST https://api.exa.ai/search with highlights.
• Retries: Exa 3 × 6 s, DeepSeek 2 × 8 s, 300 ms backoff, 25 s overall budget.
• Results are de-duplicated on a normalized URL (tracking params, hash and trailing slash removed). Hits found by both engines rank first, then by RRF score; the richer title/snippet is kept.
• Snippets are clipped to 800 characters to keep phone-sized contexts usable.

Compatibility

• DSH 0.2.1-rc.4 and the matching @deepseek-ai/schemastery (peer).
• Node.js >= 20 (native fetch, AbortSignal.any).
• Zero runtime dependencies beyond the host's schemastery.

Tests

node test/run.mjs            # merge unit tests + live legs + missing-key behavior
node test/live-takeover.mjs  # mounts on a real Cordis ctx.web and checks slot takeover

The live tests read keys from HYBRID_PATCH (defaults to a local profile patch) or DEEPSEEK_API_KEY / EXA_API_KEY.

Related

• deepseek-exa-search-combine-mcp — the same two engines combined as a standalone stdio MCP server, for hosts that cannot take a ctx.web provider.

License

MIT