Before you file

• I searched existing issues and did not find an existing proposal.
• I read the contributing guide and will wait for the Graphiti team to discuss this issue and add rfc-approved before opening a feature PR.

Affected component

Documentation or examples

Size

Unsure

User problem

Graphiti can ingest external information, but the examples do not show how to discover public web sources and turn fetched content into episodes with source provenance. Developers building research assistants need to assemble that path themselves.

Desired outcome

A runnable, opt-in example that searches the web, fetches a selected page, ingests its content as a Graphiti episode, and queries the resulting graph. Parallel Search MCP provides free web search and page fetching without a Parallel account or API key; anonymous search uses Fast mode.

Proposed design

Add an example under examples/ using a Python MCP client over streamable HTTP to https://search.parallel.ai/mcp. Discover and call web_search, show candidate URLs, and fetch a user-selected page with web_fetch. Pass bounded fetched text to Graphiti.add_episode, retain the source URL and retrieval time, and demonstrate Graphiti.search over the ingested episode. The retrieval time would represent when the content was observed, not when its claims became true.

Keep dependencies and setup local to the example, with explicit timeouts, error handling, and no graph reset. Validate the anonymous search/fetch path separately from ingestion, then test the documented end-to-end path with a supported database and Graphiti model configuration. No autonomous agent loop is proposed.

Alternatives considered

Users can manually fetch pages and adapt the quickstart, or build their own web-search client. A standalone example would document the complete MCP-to-episode path without adding a search provider to Graphiti's graph retrieval pipeline.

Compatibility and operational impact

No changes to core APIs, graph search, MCP server tools, database defaults, or saved configuration. Only the web search/fetch step is keyless; Graphiti ingestion still needs its existing database and LLM/embedding setup. Anonymous MCP usage has lower rate limits. Fetching and ingestion would be bounded and explicitly requested.

Would this example fit the project, and is this the right scope to discuss for approval?

I work at Parallel.