doc-search

Doc Index MCP — Usage Guide

Key Concept: Boundaries

Documents are chunked (~256 tokens) and embedded. The server also detects boundaries — structural markers (chapters, sections, subsections, pages) forming a hierarchy. Each chunk belongs to a boundary. Siblings are boundaries sharing the same parent.

Search Patterns

Basic search — returns chunk fragments:

doc_search(query="authentication flow")

Full section around a match — use expand_to_boundary:

doc_search(query="error handling", expand_to_boundary="section", top_k=2, max_return_tokens=8192)

Full chapter — lower top_k, raise budget:

doc_search(query="data pipeline", expand_to_boundary="chapter", top_k=1, max_return_tokens=16384)

Sibling sections — see all peers under same parent:

doc_search(query="auth", expand_to_boundary="section", include_siblings=true, max_return_tokens=16384)

Structure-first retrieval

When you already know roughly where the answer lives, skip search entirely. Read the outline, then fetch that part directly — cheaper and more complete than guessing at a query:

doc_toc(source_name="manual", max_depth=2)
doc_get_content(source_name="manual", chapter="3", max_return_tokens=8192)

doc_get_content takes exactly one locator: boundary_id, chapter, section, or pages.

When to Use What

Need	Approach
Quick fact	Default search, top_k=3
Full context around match	expand_to_boundary="section"
Whole topic area	expand_to_boundary="chapter", top_k=1
What else is in same chapter	include_siblings=true + expansion
Exact phrase	Just search — exact matches auto-boost
Know the structure already	doc_toc then doc_get_content
Specific table data	list_tables then extract_table
One-off read, no reuse	read_document (skips indexing)

Critical: Token Budget

• Default 4096 = ~16 chunks, but only 1-2 sections
• With section expansion: use 8192+
• With chapter expansion: use 16384+
• With siblings: use 16384+
• Lower top_k when expanding (1-3)