Skip to main content
Web Search RAG augments chat responses with real-time web search results. When enabled, the system searches the web before each LLM call and injects the results as additional context alongside document RAG citations.

​
Admin Configuration

Configure in Admin → Sources → Web Search:

​
Chat-Level Configuration

Chat owners configure web search per chat in Chat Settings → Web Search:

​
End-User Sidebar Toggles

Chat owners can optionally allow end users to control web search features from the sidebar at runtime:

​
Search Providers

All providers are configured in Admin → Tools ([#/admin/tool]) with their respective API keys.

​
Pinned Pages

Pinned pages are specific URLs that are always fetched on every message, regardless of the user’s query. They serve as authoritative reference material:
  • Each URL is fetched in parallel (10s timeout, 500KB limit per page)
  • HTML is stripped to plaintext and truncated to ~4,000 characters
  • Injected into the system prompt as [pinnedN] references with priority instruction
  • Appear in citations alongside regular web search results
  • Useful for grounding a chat to specific documentation, policies, or reference pages

​
How It Works

  1. Before each LLM call, the orchestrator checks if web search is enabled for the chat
  2. Web search runs against the configured provider with domain filters applied
  3. Pinned pages are fetched in parallel (independent of the search query)
  4. Results are formatted as [webN] and [pinnedN] context blocks and injected into the message history
  5. The LLM sees both document RAG results and web results, with instructions to cite sources inline
  6. Citations from web search and pinned pages are streamed to the client alongside document citations

​
Server-Side Default Merging

When a chat has web search enabled but no explicit provider selected:
  1. The server reads appSettings.defaultWebSearchProvider and applies it
  2. defaultWebSearchConfig (maxResults, includeAnswer) is merged — chat-level overrides take precedence
  3. This happens in ai.ts before the orchestrator is invoked