Research Paper Search (Exa)

Find academic papers and preprints via Exa's research-paper index (arXiv and beyond). Calls the Exa REST API through a local script; no MCP server needed, only an EXA_API_KEY.

Setup (once)

bash
export EXA_API_KEY=your-key            # or EXA_API_KEY=... in a .env at the repo root

Get a key at https://dashboard.exa.ai/api-keys. Shared details: exa-native-base.

Run it

bash
python scripts/research_paper_search.py "<topic / method>" [-n N] [--json]

Examples:

bash
python scripts/research_paper_search.py "sparse attention mechanisms for long context transformers" -n 12
python scripts/research_paper_search.py "diffusion models comprehensive survey review" -n 10
python scripts/research_paper_search.py "Geoffrey Hinton forward-forward algorithm" -n 5

Defaults to category=research paper.

Query patterns

bash
# By topic
python scripts/research_paper_search.py "category:research paper retrieval augmented generation robustness" -n 12

# Survey / review papers (best entry point into a field)
python scripts/research_paper_search.py "category:research paper <topic> comprehensive survey review" -n 10

# By author
python scripts/research_paper_search.py "category:research paper <author name> <topic>" -n 5

# By recency — encode the year(s) in the query (don't rely on date filters alone)
python scripts/research_paper_search.py "category:research paper large language model agents 2025 2026" -n 15

To find seminal papers: search for survey/review papers first, then deep-read them to extract the foundational references they cite.

Token isolation (for a literature review)

Decompose into sub-topics and dispatch a subagent per branch (e.g. architectures, training, evaluation, limitations). Each runs 3-5 paper searches and returns a compact table (title · authors · year · venue/arXiv id · one-line contribution). Merge, dedupe by title/arXiv id, then synthesize.

Override categories with -c

research paper (default) · pdf (when the paper is a standalone PDF) · news (coverage of a result). For code accompanying a paper, use exa-code-context.

After you get results

• Deep-read abstracts/sections: python ../exa-native-base/scripts/exa.py contents <arxiv-url> --text.
• For a narrative review, see references/synthesis.md in exa-native-base — organize by theme, cite every claim, surface disagreements.
• Deliver a table (title · authors · year · link · contribution) or a themed synthesis with citations.