@bizmyth/pi-vertex-gemini-search

Web search and research tools for the Pi coding agent, backed by Vertex AI Gemini with Google Search grounding through Pi's built-in google-vertex provider.

Forked from @nilskluewer/pi-vertex-gemini-search. This fork drops its own token handling, endpoint, and environment configuration and uses whatever Vertex AI setup Pi's chat models use.

Install

pi install npm:@bizmyth/pi-vertex-gemini-search

Restart Pi after installation. Review extensions before installing them; Pi extensions run with your full user permissions.

Tools

• web_search: quick fact-check or verification. Returns a concise answer and source URLs.
• web_research: in-depth research on complex topics. Returns a structured answer and source URLs.

Both tools:

• Ground answers with Google Search and read URLs passed in the query via Gemini url_context.
• Resolve vertexaisearch.cloud.google.com redirect links to their destination URLs.
• Return the answer and source URLs to the model, and structured results (answer, sources, searchQueries, model, location, serviceTier, costUsd) to codemode scripts.
• Stream the answer into the transcript while Gemini writes it, then show a model/cost/pricing-tier summary with a collapsed preview; expand the result for the full answer and sources.
• Report their token and grounding cost as tool usage, so searches count toward the session cost.

Vertex AI setup

The tools need Pi's google-vertex provider to be configured, exactly as for Vertex AI chat models. See Google Vertex AI in Pi's provider docs. For example, with Application Default Credentials:

export GOOGLE_CLOUD_PROJECT=your-project
export GOOGLE_CLOUD_LOCATION=global
gcloud auth application-default login

A Vertex AI API key (GOOGLE_CLOUD_API_KEY) or a credential stored with /login google-vertex works too. Searches run in that project and location, and fail with Pi's auth error when Vertex AI is not configured.

The current Gemini models are served from the global location.

Configuration

• PI_VERTEX_SEARCH_MODEL: the google-vertex model to search with. Default: gemini-3.8-flash. It must be in Pi's model catalog (pi --list-models google-vertex).

Pi prices tokens from its model catalog. Google Search grounding is estimated at $14 per 1,000 queries.

Pricing tier

/search-pricing            # show the current preference
/search-pricing standard   # default
/search-pricing flex       # Flex PayGo, falling back to standard

Standard is the default because Flex requests can queue until they time out, and grounding dominates the cost of a search. When a Flex request fails or times out, the search is retried at standard, and later searches stay on standard until /search-pricing flex is run again. The preference does not persist across Pi sessions.

Development

bun install
bun run check

bun run check runs Biome, TypeScript 7 (tsc, the Go-based compiler), and Bun tests.

Releasing

Trusted publishing must already be configured for publish.yml. To release a new version:

bun pm version patch # or: minor / major
git push origin main --follow-tags
gh release create "$(git describe --tags --exact-match)" --generate-notes

Publishing the GitHub release runs the checks and publishes the matching package version to npm using OIDC. No long-lived NPM_TOKEN is required.

License

MIT