Literature Search

Literature Search

Systematic, multi-engine academic paper search.

When to Use

• User asks "find papers about X"
• User needs related work for a new project
• User wants to know the state of the art on a topic
• User asks for papers from a specific venue/author/year

Engine Selection

Choose engines based on the search goal:

Workflow

Step 1: Understand the Query

Before searching, clarify:

• Scope: Broad survey vs. specific subtopic
• Recency: All time vs. last N years vs. latest only
• Venue preference: Top-tier only? Specific conference?
• Quantity: Top 5 vs. comprehensive survey
• Depth: Quick list vs. deep research with synthesis

Step 2: Select Search Strategy

Quick search (single engine): For simple, well-defined queries. Use Semantic Scholar or arXiv directly.

Multi-engine search (2-3 engines in parallel): For broader topics. Run engines simultaneously, deduplicate results.

Deep research (Gemini): For complex, multi-faceted research questions. Gemini deep research mode synthesizes across many sources and provides a structured analysis with citations. Use this when:

• The question spans multiple subfields
• You need synthesis, not just a list of papers
• The user explicitly asks for "deep research" or "comprehensive survey"

Step 3: Execute Search

# Semantic Scholar — paper metadata, citations, author search
# Free, no API key needed
node scripts/search/semantic-scholar.mjs "query" -n 20

# arXiv — latest preprints, category filtering
# Free, no API key needed
node scripts/search/arxiv.mjs "query" -n 15 --sort submittedDate --cat cs.CL

# Tavily — general web search, AI-optimized
node scripts/search/search.mjs "query site:arxiv.org OR site:aclanthology.org" -n 10
node scripts/search/search.mjs "query" --deep  # deeper search mode

# Exa — semantic search, finding similar content
mcporter call 'exa.web_search_exa(query: "query", numResults: 10)'

# Gemini — deep research (for complex questions)
# Use gemini-3.1-pro model with web search grounding
# Prompt: "Survey the recent literature on [topic]. Identify key papers,
#          main approaches, and open problems. Cite specific papers."

# AMiner — Chinese academic community
# Uses AMINER_API_KEY

Step 4: Deduplicate & Rank

Merge results across engines:

1. Deduplicate by title similarity (fuzzy match, >90% = same paper)
2. Rank by: citation count × recency × venue tier × relevance
3. Flag if a paper appears in multiple engines (higher confidence)

Step 5: Present Results

Format as a ranked list with key metadata:

1. **[Title]** (Venue Year, Citations: N)
   Authors: [First author] et al.
   TL;DR: [1 sentence]
   Why relevant: [connection to user's query]

2. ...

Step 6: Deep Dive (Optional)

If user wants to go deeper on any paper:

• Switch to paper-reading skill
• Or add to Zotero reading queue (use zotero-management skill)

Search Tips

• Use specific terminology: "multi-agent reinforcement learning" > "MARL" > "agents working together"
• Combine with venue filter: Adding venue:ACL or category cs.CL dramatically improves precision
• Check citation chains: A highly-cited paper's references and citers are often gold
• Cross-lingual: For Chinese papers, try both English and Chinese queries
• Date filter for SOTA: Use --sort submittedDate on arXiv to find the latest approaches
• Gemini for synthesis: When you need to understand a field (not just list papers), use Gemini deep research to get a narrative overview first, then drill into specific papers

Quality Signals

When ranking, weight these signals:

• Citation count: High for established work, less meaningful for papers < 6 months old
• Venue tier: ACL/EMNLP/NeurIPS/ICML > workshops > arXiv-only
• Author reputation: Check if senior authors are established in the field
• Code availability: Papers with code are more verifiable
• Reproducibility: Clear methodology sections and experimental details