using-local-library-mcp

Using local-library MCP

Core principle

When a citekey or library-adjacent cue appears, the local library is the authoritative source. Reach for MCP tools before answering from training — the specific document in the user's corpus supersedes general knowledge of the author or work.

Triggers — reach for library tools when you see

• Explicit citekeys: @Name2023, Name2023, Name et al. 2023
• Argument questions: "What does X argue / say / claim about ...", "What is X's position on ..."
• Library references: "In my library", "from my notes", "my corpus", "this document"
• Citation requests: "Cite this", "ground this", "source for this", "evidence from"
• Named works: Any paper, book, chapter, or author name when plausible as a library item (distinguishing ambiguous names from general topics)

Tool disambiguation

Intent	Tool
Get metadata for a known citekey	show_document @Name2023
Search by topic across the corpus	search_library "topic phrase"
Browse or filter the corpus	list_documents (supports --status, --year, --author-contains, etc.)
Read a document's extracted text	get_document_text @Name2023 (note: >50 chunks triggers preview mode; follow up with range calls)

Procedure — unmarked citekey in conversation

1. User mentions @Name2023 (or unmarked but plausible citekey).
2. Call show_document @Name2023 first.
3. If response starts with Error: No document found matching..., use the Did you mean @...? suggestion.
4. If response starts with ⚠ USER-FACING ERROR:, surface the infrastructure problem — do not silently fall back to training.
5. For engagement with the argument, pull evidence via get_document_text @Name2023 or targeted search_library calls before writing.

Reading the tool output

• search_library: Starts with ### Results (N chunks from M documents); each result shows @citekey | score: X | *Title*, Section:, and Chunk index labels for follow-up.
• show_document: Markdown fields: **Title:**, **Authors:**, **Date:**, **Status:**, **Embedding status:**, **Chunks:**, **Original path:** (filesystem path for PDF fallback).
• get_document_text: Short doc (≤50 chunks) = full text in code block; long doc without range = preview + ### Sections outline; long doc with range = Chunks S–E of N label.

Rationalization table — excuses to reject

Excuse	Reality
"I know this author from training; I can just summarize"	The library has their specific document. One show_document call is cheaper than being wrong about what this version says.
"The citekey is just a label; the real question is topical"	If the user wrote a citekey, they mean that document, not a general topic.
"search_library returned nothing, so the library doesn't have it"	Try show_document @<citekey> directly. Search can miss via embedding or exact-match issues; metadata lookup is authoritative.
"Listing all documents would be tedious"	list_documents supports filters (year, author, title, citekey prefix). Use them.
"I see the citekey but I recognize the work — I'll answer from memory"	Training knowledge is a guess. The document is on disk and one call away. Answer from the disk.
"The user probably knows their own paper — I don't need to verify"	Do verify. The user asked you to engage with their corpus, not with your assumptions about it.

Red flags — stop

• Writing engagement with @Name2023 without invoking a library tool
• Paraphrasing what the author "argues" without quoted evidence from get_document_text
• Asserting specific numerics (BLEU scores, dataset sizes, year counts, etc.) without confirming them in the source
• Treating Error: No document found... as a reason to proceed from training instead of offering the fuzzy-match suggestion
• Treating ⚠ USER-FACING ERROR: as a signal to silently fall back to memory