Astro arXiv Search

Retrieve astronomy and astrophysics papers from the user's arXiv mirror API. Use when Codex needs to list papers from a given arXiv day, find papers related to an astronomy topic over a recent period, find papers similar to a given arXiv paper, or fetch details for a specific arXiv ID. Prioritize this API whenever the user wants astronomy paper discovery or lookup; if the API is unavailable or unusable, fall back to other available search methods. Default to an astronomy-researcher audience and interpret ambiguous topics in an astronomy context unless the user says otherwise.

Overview

Use the user's API instead of general web search for astronomy paper discovery. Prefer lightweight listing first, then fetch more detail only for the papers the user actually cares about. Treat this API as the primary retrieval method for astronomy paper requests and only fall back to other methods if the API is unavailable, errors out, or clearly cannot answer the request.

Choose The Endpoint

• Use /ai/by-date to answer "what papers appeared on arXiv that day?", "what astro-ph papers were posted today/yesterday/on a given day?", or navigation requests such as "what was on the previous update day?".
• Use recommend with topic=... to answer "what papers were published in this area over the last month/week/year?", "were there any papers on this topic today?", or "show me the top papers related to this topic".
• Use recommend with arxiv_id=... and period=500 to answer "find papers similar to this one", "what are the related papers for this article?", or "recommend papers similar to 2306.15611".
• Use article/{arxiv_id} to answer "show me this paper", "give me the abstract/authors/categories", or any follow-up on a specific paper ID.

Read references/api.md only when exact parameter names, response fields, or example mappings are needed.

Retrieval Priority

• For astronomy and astrophysics paper discovery, use this API first.
• If the API is unavailable, repeatedly failing, or missing the capability needed for the request, fall back to other search methods available in the environment.
• When falling back, prefer structured or official paper sources before generic web search when those options are available.
• When falling back, keep the same response discipline: use the user's language by default, report exact dates when relevant, and avoid hallucinating unavailable metadata.

Handle Dates Carefully

• When the user asks for "today", "today's arXiv", or "the latest update day", resolve the latest indexed arXiv date by calling /ai/by-date with date omitted.
• Always report the exact resolved date from the API response. Do not equate the user's local "today" with the latest indexed arXiv date unless they match.
• When the user asks for a specific day, pass YYYY-MM-DD exactly.
• When the user asks for topic-related papers on one specific day, first resolve the target date if needed, then call recommend with period=0 and that exact date.
• When the user asks for no date restriction, all available dates, or an effectively unbounded recommendation query, use period=9999.
• Never send a period value above 9999. If the user asks for a larger value, cap it at 9999 and state that the API maximum is 9999.
• Use previous_date and after_date from /ai/by-date when the user asks to move backward or forward by arXiv update day instead of calendar day.

Normalize arXiv IDs

• Keep the full versioned ID such as 2603.08470v1 when using article/{arxiv_id}.
• Strip any trailing version suffix such as v1, v2, or v12 when using similar-paper recommendation by ID.
• Example: if the user gives 2306.15611v2, call the similar-paper endpoint with arxiv_id=2306.15611.

Control Context Size

• Default to only_title=true for /ai/by-date.
• Treat categories as lightweight metadata that is safe to keep in normal list responses.
• Avoid requesting the full day listing with abstracts and authors unless the user explicitly wants exhaustive details. The full payload can consume roughly 50k-100k context.
• If the user wants more detail after a title-only listing, fetch details only for the selected paper IDs with article/{arxiv_id}.
• For topic search, keep the default limit=30 unless the user asks for a smaller or larger shortlist.
• For topic search with no date restriction, use period=9999.
• For similar-paper search by ID, use the API shape the user provided: recommend?period=500&arxiv_id={bare_id} unless they later specify another supported variant.
• If the user explicitly asks for similar papers without any date restriction, use period=9999 instead of the default 500.

Respond For Astronomy Researchers

• Assume the audience is astronomy researchers unless the user specifies another audience.
• Prefer astronomy interpretations for ambiguous topics such as stars, galaxies, dark matter, planets, cosmology, and similar domain terms.
• Use categories to expose or filter subfields such as astro-ph.EP, astro-ph.GA, astro-ph.CO, and astro-ph.IM when the response includes them.
• When using recommend, prioritize papers with higher similarity scores and surface the strongest matches first.
• Summarize results in researcher-friendly language: surface the most relevant titles first, then mention notable subtopics, methods, or objects if they are obvious from the title or abstract.
• If the user asks whether papers exist on a topic, answer the existence question first, then list the strongest matches.

Report Results Clearly

• Always include the exact date used, especially for "today/latest" requests.
• For day listings, include the API's count and optionally mention previous_date or after_date when it helps the next step.
• For topic results, include the search window (period or exact day), the topic string, and the number of returned papers.
• For similar-paper results, include the source arXiv ID used for retrieval and note that the recommendation query used the bare ID without version suffix.
• For recommend results, use similarity as the default ranking signal and recommend the highest-similarity papers first.
• Include categories when they help the user judge relevance or distinguish neighboring astro-ph subfields.
• Unless the user explicitly asks for another language, present paper information in the same language the user is currently using.
• Keep original paper metadata such as title, arxiv_id, and categories as returned by the source unless the user explicitly asks for translation or rewriting; localize the surrounding explanation and summary instead.
• For single-paper lookups, include at least title, authors, abstract summary, and categories when available.
• If the API returns no results or an unexpected shape, state that directly and avoid hallucinating missing papers.

Common Request Patterns

• "Check whether there were any stellar papers on arXiv today" Resolve the latest available arXiv date, then call recommend with topic=stars, period=0, and the resolved date.
• "What black hole papers appeared in the last month?" Call recommend with topic=black holes and period=30.
• "What galaxy evolution papers are relevant across all available dates?" Call recommend with topic=galaxy evolution and period=9999.
• "Show me stellar papers from the last 20000 days" Cap the request and call recommend with topic=stars and period=9999.
• "What astro-ph papers appeared on 2026-03-12?" Call /ai/by-date?date=2026-03-12&only_title=true.
• "Recommend papers similar to 2306.15611v2" Strip the version suffix and call recommend?period=500&arxiv_id=2306.15611.
• "Show me 2603.08470v1" Call article/2603.08470v1.