README.md

Web Search Environment

A web search environment that gives the agent search and optional web scraping tools. Supports Serper and Google Custom Search providers.

Setup

Set API credentials as environment variables depending on your search provider:

# Serper (default)
export SERPER_API_KEY="your-key"

# Google Custom Search
export GOOGLE_API_KEY="your-key"
export GOOGLE_CSE_ID="your-cse-id"

Usage

from strands_env.environments.web_search import WebSearchEnv

# Search only (default)
env = WebSearchEnv(model_factory=model_factory)

# Search + scrape
env = WebSearchEnv(model_factory=model_factory, scrape_enabled=True)

# Search + scrape with LLM summarization
env = WebSearchEnv(
    model_factory=model_factory,
    scrape_enabled=True,
    summarizer_model_factory=summarizer_factory,
)

result = await env.rollout(task)
await env.cleanup()  # close shared HTTP sessions once, after ALL rollouts (sessions are reused across episodes)

Tools

Depends on configuration:

Config	Tools
Default	serper_search
search_provider="google"	google_search
scrape_enabled=True	search + scrape
scrape_enabled=True + summarizer_model_factory	search + scrape_and_summarize

Configuration

Field	Default	Meaning
search_provider	"serper"	"serper" or "google"
search_timeout	10	Seconds per search API call
blocked_domains	—	Appended to queries as -site: exclusions
scrape_enabled	False	Expose the scrape tool
scrape_timeout	50	Seconds per page fetch
scrape_token_budget	20000	Max page tokens kept for summarization

Base knobs come from EnvironmentConfig. Non-serializable named args: search_concurrency / scrape_concurrency (a shared Semaphore = one budget across envs, an int = per-env cap), summarizer_model_factory (without it the scrape tool returns raw page content).

Reward

No built-in reward function. Supply a custom reward_fn.

System Prompt

The agent is instructed to search the web, optionally scrape pages for detail, and synthesize findings into a clear, sourced answer.