Server-side result reduction (projection, pagination, summary-only) for search tools

[sajeetharan]

Motivation

Cosmos DB search results can be large (thousands of rows, embeddings, hybrid-search hits). Under direct tool calling, the full result set flows through the model; under Code Mode (Anthropic post), the client would prefer to receive only what it will actually project. Reducing at the server saves tokens for every client, not just Code-Mode clients.

Scope — extend the following tools

• text_search
• vector_search
• hybrid_search
• get_recent_documents

New optional parameters

• projection: string[] — dotted field paths to include (maps to SQL SELECT)
• maxResults: int — hard cap (existing default preserved for back-compat)
• continuationToken: string — for pagination
• summaryOnly: bool — return { count, ids[], continuationToken } only

Acceptance criteria

• New params surfaced in each tool's JSON schema
• Cosmos SQL generated safely (parameterized, reuse McpToolRequestValidator)
• Unit tests for each new mode
• docs/USE-CASES.md updated with examples

Related

Part of the "Code Mode" support umbrella.