Serper

Serper

Google search via Serper API. Fetches results AND reads the actual web pages to extract clean full-text content via trafilatura. Not just snippets — full article text.

How It Works

1. Serper API call — fast Google search, returns result URLs instantly
2. Concurrent page scraping — all result pages are fetched and extracted in parallel using trafilatura with a 3-second timeout per page
3. Streamed output — results print one at a time as each page finishes

Each invocation gives you 5 results (default mode) or up to 6 results (current mode), each with full page content. This is already a lot of information.

Query Discipline

Craft ONE good search query. That is almost always enough.

Each call returns multiple results with full page text — you get broad coverage from a single query. Do not run multiple searches to "explore" a topic. One well-chosen query with the right mode covers it.

At most two calls if the user's request genuinely spans two distinct topics (e.g. "compare X vs Y" where X and Y need separate searches, or one default + one current call for different aspects). Never more than two.

Do NOT:

• Run the same query with different wording to "get more results"
• Run sequential searches to "dig deeper" — the full page content is already deep
• Run one search to find something, then another to follow up — read the content you already have

When to Use This Skill

Use serper when:

• Any question that needs current, factual information from the web
• Research topics that need full article content, not just snippets
• News and current events
• Product info, prices, comparisons, reviews
• Technical docs, how-to guides
• Anything where reading the actual page matters

Do NOT use this skill for:

• Questions you can answer from your training data
• Pure math, code execution, creative writing
• Greetings, chitchat

IMPORTANT: This skill already fetches and extracts full page content. Do NOT use web_fetch, WebFetch, or any other URL-fetching tool on the URLs returned by this skill. The content is already included in the output.

Two Search Modes

There are exactly two modes. Pick the right one based on the query:

default — General search (all-time)

• All-time Google web search, 5 results, each enriched with full page content
• Use for: general questions, research, how-to, evergreen topics, product info, technical docs, comparisons, tutorials, anything NOT time-sensitive

current — News and recent info

• Past-week Google web search (3 results) + Google News (3 results), each enriched with full page content
• Use for: news, current events, recent developments, breaking news, announcements, anything time-sensitive

Mode Selection Guide

Locale (REQUIRED for non-English queries)

Default is global — no country filter, English results. This ONLY works for English queries.

You MUST ALWAYS set --gl and --hl when ANY of these are true:

• The user's message is in a non-English language
• The search query you construct is in a non-English language
• The user mentions a specific country, city, or region
• The user asks for local results (prices, news, stores, etc.) in a non-English context

If the user writes in German, you MUST pass --gl de --hl de. No exceptions.

Rule of thumb: If the query string contains non-English words, set --gl and --hl to match that language.

How to Invoke

python3 scripts/search.py -q "QUERY" [--mode MODE] [--gl COUNTRY] [--hl LANG]

Examples

# English, general research
python3 scripts/search.py -q "how does HTTPS work"

# English, time-sensitive
python3 scripts/search.py -q "OpenAI latest announcements" --mode current

# German query — set locale + current mode for news/prices
python3 scripts/search.py -q "aktuelle Preise iPhone" --mode current --gl de --hl de

# German news
python3 scripts/search.py -q "Nachrichten aus Berlin" --mode current --gl de --hl de

# French product research
python3 scripts/search.py -q "meilleur smartphone 2026" --gl fr --hl fr

Output Format

The output is a streamed JSON array — elements print one at a time as each page is scraped:

[{"query": "...", "mode": "default", "locale": {"gl": "world", "hl": "en"}, "results": [{"title": "...", "url": "...", "source": "web"}, ...]}
,{"title": "...", "url": "...", "source": "web", "content": "Full extracted page text..."}
,{"title": "...", "url": "...", "source": "news", "date": "2 hours ago", "content": "Full article text..."}
]

The first element is search metadata. Each following element contains a result with full extracted content.

Result fields:

• title — page title
• url — source URL
• source — "web", "news", or "knowledge_graph"
• content — full extracted page text (falls back to search snippet if extraction fails)
• date — present when available (news results always, web results sometimes)

CLI Reference