Citation Assistant

Workflow

1. Identify Target Section

When invoked:

• If section argument provided: Read that section from the LaTeX file
• If no argument: Use AskUserQuestion to ask "Which section needs citations?"
  • Offer sections found in the document (scan for \section{...})

2. Identify Citation Gaps

Read the section and identify:

• Claims without citations
• Broad claims with insufficient citations (single citation for major claims)
• Methods/techniques missing original paper references
• Comparisons to prior work without references

List gaps to user before searching - Show each claim and why it needs a citation.

3. Search for Papers

For each gap, use multiple search methods to ensure broad coverage:

1. Check existing bibliography first - Search the project's .bib file(s)

2. Search using ALL available tools (not just one):

   • Exa MCP (if available) - semantic search, good for conceptual queries
   • Semantic Scholar API - comprehensive academic database
   • Web search - catches recent preprints and non-indexed papers

   Run searches in parallel when possible. Different tools have different coverage - a paper missing from one source may appear in another.

3. Aggregate and deduplicate results - Merge results from all sources

4. Verify each paper exists in 2+ sources before citing

5. Fetch BibTeX via DOI - Never generate from memory

4. Add Citations

• Append new entries to .bib file
• Insert citation commands matching the document's style (\cite{}, \citep{}, \citet{}, etc.)
• Report what was added

Citation Rules

Never hallucinate citations. If verification fails at any step, mark the citation as needing manual verification.

BibTeX Format

@article{author2024keyword,
  title={Title with {Acronyms} Protected},
  author={Last, First and Last2, First2},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2024}
}

@inproceedings{author2024conf,
  title={Title Here},
  author={Last, First},
  booktitle={Proceedings of NeurIPS},
  year={2024}
}

• Key format: author_year_keyword
• Protect acronyms with braces: {LLM}, {GPU}, {BERT}
• Use and between authors

Python Environment

When running Python code for citation lookups:

1. Create a virtual environment if needed: uv venv
2. Activate it: source .venv/bin/activate
3. Install dependencies: uv add semanticscholar requests

Never use the system Python directly.

APIs

See references/apis.md for Semantic Scholar and DOI lookup code.

Output

For each citation added, report:

1. The claim being supported
2. Paper title and authors
3. Why this paper is relevant
4. The BibTeX key added