Web search

Web search routing, configuration, session search, and when to use web search vs. paper search.

Feynman’s web search tool retrieves current information from the web during research workflows. It supports multiple simultaneous queries, domain filtering, recency filtering, and full-page content retrieval. The Researcher agent uses web search alongside AlphaXiv to gather evidence from non-academic sources.

Default behavior

The default path is zero-config Gemini grounding via a signed-in Chromium profile. No API keys are required. This works on macOS and Linux where a Chromium-based browser is installed and signed in to a Google account.

For headless environments, CI pipelines, or servers without a browser, configure an explicit API key for Perplexity or Gemini.

Routing modes

Feynman supports three web search backends:

Mode	Description
auto	Prefer Perplexity when a key is configured, fall back to Gemini.
perplexity	Force Perplexity Sonar for all web searches.
gemini	Force Gemini grounding (zero-config default).

Configuration

Check the current search configuration:

feynman search status

This prints the active route, configured providers, browser profile, and config file path.

Edit ~/.feynman/web-search.json to configure the backend:

{
  "route": "auto",
  "perplexityApiKey": "pplx-...",
  "geminiApiKey": "AIza..."
}

Set route to auto, perplexity, or gemini. When using auto, Feynman prefers Perplexity if a key is present, then falls back to Gemini.

Search features

The web search tool exposes several capabilities that the Researcher uses automatically:

• Multiple queries — Send 2–4 varied-angle queries simultaneously for broader topic coverage. The Researcher generates queries from different angles to catch sources that use different terminology.
• Domain filtering — Restrict results to specific domains such as arxiv.org, github.com, or nature.com using domainFilter.
• Recency filtering — Filter results by date using recencyFilter, useful for fast-moving topics where only recent work matters.
• Full content retrieval — Fetch complete page content for the most important results using includeContent: true, rather than relying on snippets alone.

Web search vs. paper search

Use web search and paper search for different kinds of questions:

Use web search for

• Current products, companies, and markets
• Model availability and pricing
• Software releases and documentation
• Regulations, news, and recent events
• Anything phrased as “latest”, “current”, “recent”, or “today”
• Engineering topics with official docs and code repositories

Use paper search (AlphaXiv) for

• Academic literature and foundational research
• Survey-style coverage of a field
• Specific paper reading and Q&A
• Citation chains and related work discovery
• Benchmarks described in published papers

Never answer a latest/current question from paper search alone. Academic databases lag current events. For AI model or product claims, prefer official vendor pages and recent web sources over old papers.

For mixed topics — such as a new model that has both a launch announcement and a companion paper — the Researcher combines both channels: web sources for current reality and paper sources for the underlying research.

Session search

Session search lets you recover prior Feynman work from stored session transcripts. Every Feynman session is persisted to disk at ~/.feynman/sessions/.

Session search is an optional package. Install it with:

feynman packages install session-search

Once installed, the /search slash command becomes available in all future sessions:

/search transformer scaling laws

You can also reference prior work naturally in conversation. Feynman invokes session search automatically when you mention previous research or ask to continue earlier work — for example, “pick up where I left off on protein folding.”

What session search indexes:

• Full session transcripts including your prompts and Feynman’s responses
• Tool outputs and agent results from workflows
• Generated artifacts such as drafts, reports, and comparison matrices
• Metadata like timestamps, topics, and workflow types

The search uses both keyword matching and semantic similarity. Results include the session ID, timestamp, and relevant excerpts so you can identify which session contains the information you need.

Use session search before launching a new round of research on a topic you have investigated before. It can surface prior findings, citations, and artifacts without re-running an expensive workflow.