Repository navigation
Description
I work with Parallel on developer relations. This proposal was prepared with AI assistance.
Problem
Users running main-graph-1.1.py cannot select Parallel for a public web query through this bridge. The current handler dispatches Tavily separately from graph search and passes its Markdown into the streaming or JSON chat response. Its web formatter uses each result's original URL. That is a useful existing pattern for an optional sibling.
Open WebUI supports MCP connections in its admin settings. That provides an alternative at the host level; it does not register a selectable web mode in this bridge. The proposed workflow is to enable Parallel in the bridge, select its web-only entry from the bridge's available choices, and submit an explicitly public query, such as finding Python's official documentation.
Proposed scope
Add one disabled-by-default Parallel web mode to main-graph-1.1.py, advertised through /v1/models only when enabled. The bridge would call Unified MCP's web_search and return excerpts with the original source URLs through /v1/chat/completions. The server configuration value would be https://mcp.parallel.ai/mcp; see Parallel MCP setup.
Only the query explicitly submitted to this web mode would be sent to Parallel. Graph context, earlier conversation messages, and system messages would stay outside that request. Existing graph, Tavily, comprehensive-search, and default routing would retain their behavior. The scope would include bounded timeouts, cancellation and provider-error handling, an explicit empty-result response, and a bridge-specific User-Agent.
Validation
I inspected the source at b8b23cce8b91897b27928a4883c19abfb798ae0b and checked the dispatch and response paths with Python AST inspection. There is no Parallel integration in the tracked files. In this variant, Tavily is commented out of the advertised choices and requests require initialized graph engines; the proposal does not claim to provide standalone startup without the project's graph setup.
Before a PR, validation would use a sanitized public-documentation fixture to assert exact original URLs in both JSON responses and reconstructed streaming output. It would also cover disabled/default routing, no Parallel calls from other modes, unavailable and empty results, timeout and cancellation behavior, and a live no-key Unified MCP search when available. The live request would verify the User-Agent. Full application startup and live retrieval have not been tested for this proposal.
The README requests an issue before major changes. Adding an external provider and an MCP client lifecycle seems substantial enough to discuss first. Would you welcome this scope in the graph variant, or prefer to keep external search configuration in Open WebUI?
Activity