zotero-research-lookup

---

name: zotero-research-lookup description: Research from a local Zotero library through MCP. Use a deterministic workflow: check search DB status, run semantic search, retrieve metadata + fulltext for top items, and fall back to keyword search when needed. allowed-tools: [Read, Bash] license: MIT license

Zotero Research Lookup

Overview

Use this skill when the source of truth should be the user's local Zotero library instead of web search APIs.

Primary workflow:

1. Check semantic index status and retriever mode
2. Run semantic search
3. Use matched_content (advanced_rag) or retrieve fulltext (legacy) for evidence
4. Return traceable citations (item key and Zotero URL)

When to Use

Use this skill for:

• Local literature review and evidence gathering
• Verifying claims against already curated papers
• Draft support where stable citations are required
• Finding methods, protocols, or technical details from saved papers
• Querying notes/annotations from local library context

Do not use this skill when:

• The request requires very recent papers not yet in the user's Zotero library
• Statistical data from current surveys/reports is needed (Zotero indexes what the user has saved, not the full web)
• Citation counts, journal impact factors, or h-index data are required (not available in Zotero metadata)

Scope limitation: This skill searches only the user's curated Zotero library. Coverage depends entirely on what the user has imported. If a topic returns few or no results, fall back to claude-scientific-writer:semantic-scholar-lookup to search the open Semantic Scholar corpus.

Retriever Modes

The retriever_mode field in zotero_get_search_database_status determines search behavior:

Mode	Index content	Evidence source	Quality
legacy_metadata	Title + abstract + authors	Call zotero_get_item_fulltext separately	Basic
legacy_fulltext	Metadata + extracted PDF text	Call zotero_get_item_fulltext separately	Moderate
advanced_rag	Markdown chunks (paragraph-level) with Rerank	matched_content in search result	Best

advanced_rag mode (recommended) requires:

• retriever_mode: "advanced_rag" in ~/.config/zotero-mcp/config.json
• md_root pointing to MinerU-converted Markdown files
• Index built with zotero-mcp update-db --rebuild

In advanced_rag, each result already contains matched_content—the exact chunk that matched the query, reranked by a cross-encoder. Do not call zotero_get_item_fulltext for every item; use matched_content directly and call fulltext only when deeper context is needed.

Required MCP Tools

• zotero_get_search_database_status
• zotero_semantic_search
• zotero_get_item_metadata
• zotero_get_item_fulltext (legacy modes or when deeper context needed)
• zotero_search_items (keyword fallback)

Default Parameters

• Semantic search limit: 5
• include_citation_references: true (advanced_rag returns resolved_citations per item)
• citation_max_items_per_result: 8
• Evidence snippets per item: 1-3 from matched_content (advanced_rag) or fulltext (legacy)
• Retrieval policy: semantic search first, keyword search fallback

Deterministic Workflow

Step 1: Check Index Health

Call zotero_get_search_database_status.

Inspect the response:

• retriever_mode: determines evidence strategy (see table above)
• collection_info.document_count: if 0, the index is empty
• advanced_rag.candidate_k: number of candidates before reranking (default 30)
• advanced_rag.meta_weight: metadata blending weight (default 0.70)
• reranker_status.enabled: whether FlashRank reranking is active

If document count is 0 or not initialized:

• advanced_rag mode: run zotero-mcp update-db --rebuild (requires md_root configured)
• legacy modes: run zotero-mcp update-db (add --fulltext for PDF text extraction)

Step 2: Semantic Retrieval

Call zotero_semantic_search with:

• query: user question
• limit: default 5 (or user override)
• filters: optional metadata filter dict (e.g. {"year": "2023"})
• include_citation_references: true to get resolved_citations (advanced_rag only)
• abstract_max_chars / matched_content_max_chars: optional truncation controls

If semantic search errors or returns empty, fallback to zotero_search_items using the same query.

Step 3: Evidence Collection Per Item

advanced_rag mode (preferred):

1. Extract item_key, similarity_score, matched_content directly from result
2. matched_content = the reranked chunk(s) that best matched the query—use as primary evidence
3. resolved_citations = reference list entries cited within this chunk (e.g. [1] Author et al. 2020)—use to surface additional relevant papers
4. Call zotero_get_item_metadata(item_key) for canonical citation fields
5. Call zotero_get_item_fulltext only if matched_content lacks sufficient context

legacy modes:

1. Extract item_key and similarity_score from result
2. Call zotero_get_item_metadata(item_key) for citation fields
3. Call zotero_get_item_fulltext(item_key) and extract 1-3 relevant snippets
4. If no fulltext available, keep metadata-only evidence and mark the limitation

Query Construction

Zotero semantic search uses vector embeddings—query phrasing matters. Use these patterns:

Structured format: [concept] + [aspect] + [context/field]

Query type	Example
Literature review opening	"theoretical frameworks for [topic] review"
Empirical evidence	"[intervention] effect on [outcome] [population]"
Methods/protocols	"[method name] procedure protocol [field]"
Gap identification	"limitations future research [sub-topic]"
Comparative	"[concept A] versus [concept B] [field]"
Mechanism	"mechanism [phenomenon] [field]"

Tips:

• Use field-specific terminology matching the papers in your library
• For broad topics, run multiple targeted queries per sub-theme rather than one vague query
• Apply filters (e.g. {"year": "2022"}) to scope searches by year or other metadata
• If results are poor, try synonyms or broader phrasing

Step 4: Response Assembly (Expanded)

Return output in this order:

1. Query summary (1-2 lines)
2. Ranked evidence list
3. Synthesis across items
4. Gaps/uncertainty notes

Per item include:

• Rank
• Title, authors, year, journal/venue
• item_key and similarity score
• 1-3 evidence snippets (from matched_content or fulltext)
• resolved_citations if present (list notable cited references)
• Link: zotero://select/items/<ITEM_KEY>

Synthesis should cover:

• Research consensus: what do multiple sources agree on?
• Contradictions or debates across sources
• Methodological patterns (what methods dominate?)
• Identified research gaps explicitly mentioned in the papers
• Which sources are most directly relevant to the query

Note: Zotero metadata includes journal name and year but not citation counts or impact factors. When journal quality matters, report the journal name and let the user assess.

Output Contract

Use this structure:

## Query
<user query>

## Top Evidence
### 1) <title>
- item_key: <key>
- authors: <authors>
- year: <year>
- relevance: <score>
- evidence:
  - "..."
  - "..."
- resolved_citations: [1] Author et al. 2020 — <title snippet>
- zotero_url: zotero://select/items/<key>

## Synthesis
- ...

## Limitations
- ...

Failure Handling

• Single-item failure: skip item, continue with remaining results, and report skipped key.
• No matched_content (advanced_rag): fall back to zotero_get_item_fulltext.
• No fulltext available: retain metadata result and report that only metadata-level evidence exists.
• Context budget exceeded: reduce limit first, then shorten snippet count per item.
• Semantic search returns empty or irrelevant results: fall back to claude-scientific-writer:semantic-scholar-lookup to search the open Semantic Scholar corpus with the same query.

Quick Start

Use scripts/run_lookup.sh to generate a ready-to-paste execution prompt for Claude.

Examples are in scripts/examples.md.