gosearch

gosearch

Web search and page-content extraction for Go — no API key, no SDK, no account.

gosearch queries real search engines (DuckDuckGo, Bing, Google, Yandex) by fetching their public HTML result pages and parsing them directly, and extracts the readable main content of any URL. It is built for local-first programs — LLM agents, CLI tools, self-hosted services — that cannot or will not depend on a hosted search API.

results, _ := gosearch.Search(ctx, "golang html parser", gosearch.DuckDuckGo)

page, _ := gosearch.Fetch(ctx, results[0].URL,
    gosearch.WithMarkdown(), // headings, lists, links, code fences preserved
)
fmt.Println(page.Content) // ready to feed an LLM

Contents

• Why
• Features
• Quick start
• Usage patterns
• Choosing an engine
• Options
• Reliability, honestly
• Explicit non-goals
• Real-browser rendering (gosearch/browser)
• Versioning & stability
• Documentation
• Contributing
• License

Why

Most "web search for my agent/tool" solutions assume you'll pay for and depend on a hosted search API (SerpAPI, Google Custom Search, Bing Search API…). That is a real dependency: an account, a key, a bill, a third party on your critical path. Projects built to run fully locally need none of those things — so gosearch talks to the engines' own public result pages directly, the same way a web browser would, and parses what comes back.

The core module's entire third-party dependency surface is golang.org/x/net/html. That is a deliberate architectural bet: you can drop this into almost any Go program without dragging in an SDK.

Features

• Four engines, one interface — DuckDuckGo, Bing, Google, Yandex through a single Search(ctx, query, engine, ...Option) call returning structured []Result{Title, URL, Snippet}.
• Readable content extraction — Fetch(ctx, url, ...Option) returns a page's main content with navigation, ads, scripts, and boilerplate removed — never raw HTML.
• Markdown output — WithMarkdown() renders extracted content as GitHub-flavored Markdown: headings, lists, tables, fenced code, links, and emphasis survive. Built for LLM context, where structure carries meaning.
• Ordered fallback chain — WithFallback(...) moves to the next engine only when the current one reports being blocked or challenged, instead of failing outright.
• Honest failure signaling — anti-bot interventions surface as typed sentinel errors (ErrBlocked, ErrChallenge) via errors.Is, never as a silent empty result or a parse panic on a captcha page.
• Near-duplicate collapsing — engines often list one page under several URL spellings (percent-encoded vs literal İ/I, reordered parameters); these collapse into one result, original spelling preserved.
• Opt-in freshness dates — WithDates() fills Result.Date from each engine's own metadata when provided; off by default so timeless use cases are never polluted by accident.
• Caller-side domain policy — WithBlockedDomains(...) / WithAllowedDomains(...) enforce your spam and quality rules on the result list. The library judges no site on its own.
• Resilience built in — transient-failure retries with exponential backoff, realistic browser headers, persistent cookie jar, per-host rate limiting.

Quick start

Requirements: Go 1.25+. No keys, no accounts, no config files.

go get github.com/BugraAkdemir/gosearch@latest

package main

import (
	"context"
	"fmt"
	"log"

	"github.com/BugraAkdemir/gosearch"
)

func main() {
	ctx := context.Background()

	// Search (DuckDuckGo is the most reliable default).
	results, err := gosearch.Search(ctx, "facebook", gosearch.DuckDuckGo,
		gosearch.WithMaxResults(5),
	)
	if err != nil {
		log.Fatal(err)
	}
	for _, r := range results {
		fmt.Println(r.Title, "->", r.URL)
	}

	// Extract any page's readable content.
	page, err := gosearch.Fetch(ctx, "https://en.wikipedia.org/wiki/Facebook")
	if err != nil {
		log.Fatal(err)
	}
	fmt.Println(page.Title)
	fmt.Println(page.Content)
}

Usage patterns

Resilient searches with fallback

If the primary engine blocks or challenges the request, the same query continues down the chain automatically. First success wins.

results, err := gosearch.Search(ctx, "facebook", gosearch.Google,
	gosearch.WithFallback(gosearch.Bing, gosearch.DuckDuckGo),
)

Page content as Markdown (LLM-ready)

page, err := gosearch.Fetch(ctx, url, gosearch.WithMarkdown())
// page.Content now contains "# Heading", "- list items", "[links](…)", "```code```"

The LLM-agent pipeline

The two calls compose into the standard retrieval flow:

// 1. Find candidate sources.
results, err := gosearch.Search(ctx, question, gosearch.DuckDuckGo,
	gosearch.WithFallback(gosearch.Bing),
	gosearch.WithMaxResults(5),
	gosearch.WithBlockedDomains("pinterest.com"), // your quality policy
)

// 2. Read the top hits as Markdown context.
var context strings.Builder
for _, r := range results[:min(3, len(results))] {
	page, err := gosearch.Fetch(ctx, r.URL, gosearch.WithMarkdown())
	if err != nil {
		continue // one unreadable source ≠ abort
	}
	fmt.Fprintf(&context, "## %s\n%s\n\n", page.Title, page.Content)
}
// context.String() → model input

Freshness-aware results

results, _ := gosearch.Search(ctx, q, gosearch.Bing,
	gosearch.WithDates(), // Result.Date: "1 day ago", "2026-08-20", …
)

Full option semantics: docs/API.md. Task-oriented copy-paste recipes: docs/RECIPES.md.

Choosing an engine

Anti-bot strictness differs per engine and per network — IP reputation is often the deciding factor, not client behavior. Rough expectations from a normal residential connection, based on live testing of this library:

"Best-effort heuristic" status: the DuckDuckGo parser is validated against a real captured response; the other three parsers match current observed markup and are tested against synthetic fixtures, but engines rotate their DOM without notice — treat a parse miss as a signal to capture fresh HTML, not necessarily a bug.

Options

One variadic ...Option applies to both Search and Fetch; scope notes below. Full table with semantics: docs/API.md.

Reliability, honestly

Search engines run anti-bot systems. A 200 OK does not mean success — engines serve captcha/challenge pages with status 200, which is why every response passes through block detection before parsing, and why failures are typed errors rather than empty slices. From datacenter/cloud IPs every engine may challenge or block you regardless of politeness; residential networks fare far better. When every engine in a fallback chain refuses you, Search returns an errors.Join of each engine's error so you can see exactly what happened.

Explicit non-goals

This library behaves like an ordinary visitor — and stops there. It will never solve CAPTCHAs, execute JS challenges to disguise itself, mask automation flags, or rotate identities to defeat a security control. Those cross from "look like a normal visitor" into "defeat a security control," which is out of scope on principle, not merely as a technical limitation.

Real-browser rendering (gosearch/browser)

For pages whose content only exists after JavaScript runs — which defeats plain-HTTP Fetch() by definition — an opt-in separate Go module drives a real, unmodified Chromium-family browser:

• Discovers Chrome/Edge/Chromium on the system; optionally downloads Google's official chrome-headless-shell with explicit permission; or embeds the engine at compile time via -tags gosearch_embed_engine.
• One long-lived process and one reused tab: steady-state cost is about one page's RAM, not a browser per request.
• Same honest line: it executes JavaScript; it does not solve CAPTCHAs or mask automation.

Because it lives in its own module, go get github.com/BugraAkdemir/gosearch never pulls browser dependencies into your project. See browser/README.md for installation and trade-offs.

Versioning & stability

Releases follow semantic versioning as annotated vX.Y.Z tags; the module proxy distributes them automatically. During v0.x, the public API may still evolve between minor versions — pin an exact version if that matters to you. The browser/ directory is a separate Go module with its own prefixed tags (browser/vX.Y.Z).

Documentation

Contributing

See CONTRIBUTING.md. Development conventions, verification commands, and known pitfalls live in AGENTS.md; the phased roadmap and exit criteria live in plan.md.

License

MIT