Tavily arXiv Paper Fetech Pipeline#

Orchestrate a title-to-arXiv metadata workflow for: $ARGUMENTS

Overview#

This skill turns a list of paper titles into a reproducible arXiv lookup result:

text

paper titles
-> Tavily search only
-> canonical arXiv abs URL
-> local arXiv metadata fetch
-> JSONL process log
-> markdown report

The goal is not just to answer once. The goal is to leave behind a reusable work folder that can be resumed or consumed by another workflow.

Inputs#

Extract or ask for:

• TITLE_LIST: one or more paper titles
• WORKDIR: optional working folder

If WORKDIR is omitted, default to:

• 0_docs/tavily_arxiv_paper_fetech

Accepted title formats:

• a single title
• multiple titles separated by newlines
• markdown bullet lists

Constants#

• TAVILY_ONLY = true — Tavily is the only allowed search mechanism
• SEQUENTIAL_MODE = true — process titles one by one to reduce rate limits
• NO_FALLBACK_SEARCH = true — never switch to arXiv API search, guessed arXiv URLs, or another provider
• LOW_CONTEXT_MODE = true — append one JSON line per title, then render markdown from JSONL

Work Folder Layout#

Write all outputs under WORKDIR:

ArtifactPurposeinput_titles.mdnormalized input title listpaper_fetches.jsonlone JSON line per processed titlepaper_fetch_report.mdrendered report from the JSONL

Execution Rule#

Process titles in order. Do not parallelize Tavily calls.

If Tavily rate limits:

• wait briefly
• retry Tavily only
• if it still fails, record the error and continue

Phase 0: Initialize#

1. Normalize the input title list.
2. Create WORKDIR.
3. Write the normalized title list to WORKDIR/input_titles.md.

Phase 1: Tavily resolution#

For each title:

1. Search Tavily with:

text

"<paper title>" arXiv

1. Prefer results in this order:
   • https://arxiv.org/abs/...
   • https://arxiv.org/html/...
   • https://arxiv.org/pdf/...
2. Accept a result only when the match is reliable:
   • exact title match after minor normalization
   • same title with only punctuation, Unicode, or math-formatting differences
   • arXiv page clearly shows the same paper title
3. If uncertain, record no_match instead of guessing.

Phase 2: Local arXiv fetch#

Once a reliable arXiv URL is known, run:

bash

python3 ".cursor/skills/tavily-arxiv-paper-fetech/scripts/fetch_arxiv_abs.py" "<arxiv-url>"

This returns compact JSON with:

• canonical abs URL
• arXiv id
• title
• authors
• abstract

Phase 3: Low-context JSONL logging#

For each title, immediately append one JSON line to:

• WORKDIR/paper_fetches.jsonl

Each line should include at least:

• index
• input_title
• tavily_status
• tavily_error
• arxiv_url
• fetch

Status values:

• ok
• no_match
• error

Phase 4: Render markdown report#

After all titles are processed, run:

bash

python3 ".cursor/skills/tavily-arxiv-paper-fetech/scripts/jsonl_to_paper_fetch_md.py" \
  "WORKDIR/paper_fetches.jsonl" \
  "WORKDIR/paper_fetch_report.md"

Output Format#

The rendered report should look like:

markdown

# Tavily arXiv Paper Fetech Report

## Results

### 1. Original Title
- Status: ok
- arXiv URL: https://arxiv.org/abs/xxxx.xxxxx
- arXiv ID: xxxx.xxxxx
- Resolved Title: ...
- Authors: ...
- Abstract: ...

Key Rules#

• Never fabricate a paper match.
• Never use non-Tavily search for title resolution.
• Keep all outputs inside WORKDIR.
• Prefer the local helper script over bringing full arXiv page content into context.

Utility Scripts#

• scripts/fetch_arxiv_abs.py — fetch compact metadata from a known arXiv URL
• scripts/jsonl_to_paper_fetch_md.py — render JSONL to markdown

Additional Resources#

• For copy-paste invocations and expected outputs, see examples.md

Example Invocation#

text

/tavily-arxiv-paper-fetech "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control\nOpenVLA: An Open-Source Vision-Language-Action Model --workdir 0_docs/tavily_arxiv_lookup_run_01"