Gabfssilva Scimesh Papers

Papers Search Skill

Search scientific papers using scimesh via uvx (no global installation required).

Environment Variables

Available Providers

Basic Usage

Execute commands directly with uvx scimesh:

# Search with field operators
uvx scimesh search "TITLE(transformer)"

# Plain text also works (searches in title AND abstract)
uvx scimesh search "transformers"
uvx scimesh search "attention mechanism"

Triggers

• "search papers about X"
• "papers about Y"
• "I want to read about Z"
• "recent papers on W"
• "what's new in transformers?"
• "save this paper to workspace"

Search Flow

1. Understand the intent: extract topic/query from user message
2. Refine with questions (use AskUserQuestion):
   • Year filter? (2024 / 2023-2024 / 2020-2024 / Run search)
   • Specific area? (ML / NLP / CV / Run search)
   • Preferred sources? (All / arXiv / OpenAlex / Run search)
   • IMPORTANT: Always include "Run search" as an option so the user can skip refinements
3. Connectivity check (BEFORE executing search):

   {
       "questions": [
           {
               "question": "Are you connected to institutional VPN?",
               "header": "VPN",
               "options": [
                   {"label": "Yes, VPN active", "description": "Better access to papers"},
                   {"label": "No VPN", "description": "Will use Open Access only"},
                   {"label": "Wait, connecting...", "description": "Pause for VPN connection"}
               ],
               "multiSelect": False
           }
       ]
   }
   

   • If user chooses "Wait", pause and ask again when ready
4. Execute search using Scopus syntax:

   uvx scimesh search "TITLE(transformer) AND PUBYEAR > 2020" -p arxiv,openalex -n 20
   

5. Present results in a readable format

Full Paper Reading Flow

When the user wants to read a complete paper (fullpaper):

1. Pre-Download Check

BEFORE downloading, ask about connectivity and Sci-Hub in a single batch:

{
    "questions": [
        {
            "question": "Are you connected to institutional VPN?",
            "header": "VPN",
            "options": [
                {"label": "Yes, VPN active", "description": "Institutional access enabled"},
                {"label": "No VPN", "description": "Will try Open Access only"},
                {"label": "Wait, connecting...", "description": "Pause for VPN connection"}
            ],
            "multiSelect": False
        },
        {
            "question": "Enable Sci-Hub as fallback if Open Access fails?",
            "header": "Sci-Hub",
            "options": [
                {"label": "No (Rec)", "description": "Only legal Open Access sources"},
                {"label": "Yes", "description": "Use --scihub flag (at your own risk)"}
            ],
            "multiSelect": False
        }
    ]
}

• If user chooses "Wait, connecting...", pause and ask again when ready
• If user enables Sci-Hub, add --scihub flag to download commands

2. Download the PDF

Use the scimesh download command (requires UNPAYWALL_EMAIL):

# Download via DOI (tries Open Access via Unpaywall)
uvx scimesh download "10.1234/example" -o {papers_dir}/

# With Sci-Hub fallback (use at your own risk)
uvx scimesh download "10.1234/example" --scihub -o {papers_dir}/

# Download multiple DOIs from file
uvx scimesh download -f dois.txt -o {papers_dir}/

# Search + Download directly (pipe)
uvx scimesh search "TITLE(paper)" -f json | uvx scimesh download -o {papers_dir}/

# For arXiv papers, use DOI in the format: 10.48550/arXiv.XXXX.XXXXX

PDF Caching: PDFs are automatically cached in ~/.scimesh/cache/pdfs/ - avoids re-download.

Manual alternatives:

• arXiv: https://arxiv.org/pdf/XXXX.XXXXX
• Generic DOI: Check the pdf_url field in search results

3. Folder Structure

Ask user where to save papers (if not already configured):

{
    "question": "Where should I save papers?",
    "header": "Papers dir",
    "options": [
        {"label": "Current directory", "description": "./papers/"},
        {"label": "Documents", "description": "~/Documents/papers/"},
        {"label": "Custom path", "description": "You specify"}
    ],
    "multiSelect": False
}

Save in {papers_dir}/ with the structure:

{papers_dir}/
  <year>-<first-author>-<short-title>/
    fulltext.pdf
    README.md

Note: {papers_dir} is user-defined. Examples: ./papers/, ~/Documents/papers/, ~/brain/08-papers/

Example: 2024-wang-treb/

4. Read the PDF

• Use Read tool directly on the .pdf file (native Claude Code support)
• No need for pdftotext, PyMuPDF, or other Python libs
• Claude Code extracts text + renders figures page by page

5. Systematic Reading Structure

1. Abstract - understand the main claim
2. Figures/Tables - often tell the story before the text
3. Methodology - how they did it
4. Results - specific metrics, comparisons with baselines
5. Limitations - what the authors admit (or omit)

6. Critical Verification Checklist

• Compares with relevant baselines?
• Uses standard metrics for the field?
• Tests on multiple datasets?
• Reports standard deviation/confidence intervals?
• Code available?
• Limitations section exists?

7. Common Red Flags

• Empty or superficial "Evaluation" sections
• Only one dataset
• Non-standard or invented metrics
• No quantitative comparison with existing methods
• Strong claims without proportional evidence

Save Flow

When the user asks to save a paper:

1. Get complete details (use previous search result or search by DOI/arXiv ID)

2. Create paper folder:

   mkdir -p {papers_dir}/<year>-<author>-<title>/
   

3. Download PDF (if available):

   uvx scimesh download "DOI" -o {papers_dir}/<year>-<author>-<title>/
   

4. Create README.md in the folder with the format:

   ---
   title: "Paper Title"
   authors: [Author 1, Author 2]
   year: 2024
   doi: 10.1234/example
   arxiv: 2401.12345
   tags: [paper, area, subtopic]
   saved: YYYY-MM-DD
   ---
   
   ## Summary
   
   [Generated summary - 2-3 accessible paragraphs]
   
   ## Original Abstract
   
   [Paper abstract]
   
   ## Methodology
   
   [Methodology description]
   
   ## Results
   
   [Main results and metrics]
   
   ## Critical Evaluation
   
   ### Strengths
   - [...]
   
   ### Limitations
   - [...]
   
   ## Links
   
   - [arXiv](url)
   - [PDF](./fulltext.pdf)
   - [GitHub](url) (if available)
   
   ## Connections
   
   - [[related-note-1]]
   - [[related-note-2]]
   
   ## Notes
   
   [Space for personal annotations]
   

5. Suggest connections with other workspace notes

Available Commands

Search Paper by DOI

# Search specific paper by DOI (merge from multiple providers)
uvx scimesh get "10.1038/nature14539"

# From specific providers
uvx scimesh get "10.1038/nature14539" -p openalex,semantic_scholar

# Export to BibTeX
uvx scimesh get "10.1038/nature14539" -f bibtex -o paper.bib

# arXiv paper by ID
uvx scimesh get "1706.03762" -p arxiv

Search Citations

# Papers that cite a DOI (incoming citations)
uvx scimesh citations "10.1038/nature14539" --direction in

# Papers cited by a DOI (references/outgoing)
uvx scimesh citations "10.1038/nature14539" --direction out

# Both directions
uvx scimesh citations "10.1038/nature14539" --direction both

# From specific providers (openalex, semantic_scholar, scopus)
uvx scimesh citations "10.1038/nature14539" -p semantic_scholar -n 50

# Export to JSON
uvx scimesh citations "10.1038/nature14539" -f json -o citations.json

Search with Scopus Syntax

# Search by title
uvx scimesh search "TITLE(transformer)"

# Search with multiple filters
uvx scimesh search "TITLE(transformer) AND AUTHOR(Vaswani)"

# Filter by year
uvx scimesh search "TITLE(attention) AND PUBYEAR > 2020"
uvx scimesh search "TITLE(BERT) AND PUBYEAR = 2024"

# Search in abstract
uvx scimesh search "TITLE-ABS-KEY(reinforcement learning)"

# Fulltext search
uvx scimesh search "ALL(CRISPR gene editing)"

# Combine operators
uvx scimesh search "TITLE(transformer) AND AUTHOR(Bengio) AND PUBYEAR > 2020"

# Negation
uvx scimesh search "TITLE(deep learning) AND NOT TITLE(survey)"

# Group conditions
uvx scimesh search "(TITLE(transformer) OR TITLE(attention)) AND PUBYEAR = 2024"

Command Line Options

# Limit number of results per provider
uvx scimesh search "TITLE(transformer)" -n 20

# Limit total results
uvx scimesh search "TITLE(transformer)" -t 50

# Choose providers (arxiv, openalex, scopus, semantic_scholar)
uvx scimesh search "TITLE(transformer)" -p arxiv,openalex,semantic_scholar

# Export to BibTeX
uvx scimesh search "TITLE(transformer)" -f bibtex -o papers.bib

# Export to RIS
uvx scimesh search "TITLE(transformer)" -f ris -o papers.ris

# Export to CSV
uvx scimesh search "TITLE(transformer)" -f csv -o papers.csv

# Export to JSON
uvx scimesh search "TITLE(transformer)" -f json -o papers.json

# Search + Download directly (pipe)
uvx scimesh search "TITLE(transformer)" -f json | uvx scimesh download -o papers/

# Disable deduplication
uvx scimesh search "TITLE(transformer)" --no-dedupe

Local Indexing for Fulltext

For fulltext search on providers without native support (Semantic Scholar):

# Index local PDFs
uvx scimesh index ./papers/

# Clear and re-index
uvx scimesh index ./papers/ --clear

# Now ALL() uses local index + native APIs
uvx scimesh search "ALL(attention mechanism)"

The index is stored in ~/.scimesh/fulltext.db using SQLite FTS5.

Available Search Fields

Logical Operators

Interaction Examples

User: "I want to see papers about RLHF"

Skill:

1. Ask: "Do you want to filter by year?" with options [2024, 2023-2024, 2020-2024, Run search]
2. If user chooses year, ask next filter or execute
3. Execute:

   uvx scimesh search "TITLE-ABS-KEY(RLHF) AND PUBYEAR > 2022" -n 20
   

4. Present formatted results

User: "papers about transformers in 2024"

Skill:

uvx scimesh search "TITLE(transformer) AND PUBYEAR = 2024" -n 20

User: "papers by Bengio about deep learning"

Skill:

uvx scimesh search "AUTHOR(Bengio) AND TITLE(deep learning)" -n 20

User: "save the second one to workspace"

Skill:

1. Get data from the second paper of the last search
2. Create folder: {papers_dir}/2024-author-title/
3. Download PDF: fulltext.pdf (if available via Open Access)
4. Generate README.md with complete summary
5. Suggest tags and connections

User: "read this paper for me"

Skill:

1. Use Read tool on fulltext.pdf
2. Follow systematic reading structure
3. Apply critical verification checklist
4. Identify red flags if any
5. Update README.md with complete analysis

User: "export the search to bibtex"

Skill:

uvx scimesh search "TITLE(transformer)" -f bibtex -o papers.bib

Converted and distributed by TomeVault — claim your Tome and manage your conversions.