dsh-search-first

Search First

中文

Search first, verify facts, reply last — enforced by DSH.

▶ Demo walkthrough: https://youtu.be/rGc5OCydHyY

A DeepSeek Harness (dsh) Cordis plugin that forces multi-source web search before the model answers. On agent/pre-step, when Search First is enabled, it runs Query Extraction → Unified Search Engine (AnySearch / TinyFish / Tavily) → fetch → organize → cache + runs/ persistence → inject the organized live leeds guiding verification next, then returns control to the agent loop. Final answers stay with the LLM.

Features

1. The sub-model used for query extraction is configurable.
2. Guides the host model to use standardized citations, as shown below.
3. Models and ranks multi-source search results, and selectively fetches pages back to the host model.

Installation

Prerequisites: a mounted dsh web profile (dsh ≥ 0.1.1-rc.2), Node ≥ 22.

# install from the GitHub repo:
dsh plugin --profile web add https://github.com/v587d/dsh-search-first.git

Uninstall: dsh plugin --profile web remove dsh-search-first (then restart).

Build from source

pnpm install
pnpm check      # vitest + typecheck (host + client)
pnpm build      # lib/ (host ESM + d.ts) + lib/client.js (browser bundle)
pnpm test

Quick start

[!IMPORTANT] Export at least one API key (AnySearch / Tinyfish / Tavily) before running.
Free tier is ENOUGH.

1. Restart dsh web after installing.
2. Open Settings → Search first → Search APIs → Test all connections (configure keys first via step 3, or the probe reports a key-missing result).
3. Provide at least one provider key — add it to $DSH_HOME/.credentials.yaml or export TAVILY_API_KEY / ANYSEARCH_API_KEY / TINYFISH_API_KEY in the environment. AnySearch also works anonymously (Bearer header omitted).
4. Back in a session, the Search First toggle under the input box decides per session (right-click the toggle to return to the global default).

Configuration (Settings → Search first)

License

MIT — see LICENSE.