# Searching from an AI client

Prompt patterns for querying collections over MCP.

Once the [MCP server is connected](/docs/getting-started/install/#5-connect-an-mcp-client),
ask your assistant naturally. It resolves the collection with `list_collections`,
then calls `search_collection` or `ask_collection`.

## Patterns that work

> "Search my **research** collection for React state management patterns."

> "What do my bookmarks say about database indexing?"

> "In the **gigs** collection, find remote product design roles."

> "Pull the exact wording from my **research** bookmarks about cosine vs dot-product similarity."

Naming the collection helps the client skip straight to it. If you don't, it
lists collections first and picks by name and description — so give collections
descriptions.

## What comes back

Ranked text chunks, each with its source `url` and `title`, plus an
`injection_risk` score when the passage was flagged during indexing. The
assistant synthesises from the chunks; the retrieval itself does no generation.

## Tuning

`top_k` defaults to 5. Ask for "more results" or "the top 10 passages" and a
capable client will raise it.