deepseek-web-search

DeepSeek Web Search

Searches the web server-side via DeepSeek's Responses API web_search tool and returns a synthesized answer with the source URLs the model consulted.

When to Use

• User asks about current events, recent news, or time-sensitive facts.
• You want a ready answer where someone reads pages and summarizes (vs. getting raw results).
• Chinese or English web coverage needed.
• When NOT to use: if you need a raw, structured list of search results (titles/URLs/snippets) to extract or analyze yourself, this skill is not the right fit — DeepSeek only returns its synthesis + consulted URLs.

Usage

Run the bundled script from this skill's directory (where this SKILL.md lives):

node ./search.mjs "" [maxOutputTokens]

• query (required): The search query or question.
• maxOutputTokens (optional): Cap on answer length (default 8000, max 16384).

Output

Returns a JSON object:

{
  "answer": "Synthesized answer text",
  "sources": ["https://page-the-model-opened.example", "..."],
  "query": "original query",
  "model": "deepseek-v4-flash",
  "usage": { "input_tokens": 0, "output_tokens": 0, "total_tokens": 0 },
  "engine": "deepseek-web-search"
}

Notes

• The model performs multi-round search (search → open pages → re-search) server-side, so calls take 15–70 seconds. Don't treat a delay as failure; the script allows up to 120s.
• API key comes from config.json (next to search.mjs) or the DEEPSEEK_API_KEY env var. If both are missing, the script reports an error.
• sources are the URLs the model opened while researching — a "pages consulted" list, not per-claim citations. Treat the answer as model synthesis; verify critical facts against the sources before trusting them.
• The model may notice and correct false premises in the query (e.g. "2026 Paris Olympics" → Milan-Cortina 2026 Winter Olympics).
• Always cite source URLs when presenting information to the user.