Papers

Overview

3,166,878 arXiv papers in full text, synced the day they are announced, plus PubMed Central, bioRxiv and medRxiv. Paper access and retrieval live under /arxiv/ and /pmc/; agentic search under /arxiv/agent/.

Writing queries that work matters more than any parameter. Be specific — “what compression ratio does KV cache eviction report on LongBench” beats “kv cache”. Ask for numbers if you want numbers. Chinese works directly; queries are rewritten to English technical terms for retrieval and the answer comes back in your language. Put scope limits — year, venue, category, author, institution, minimum citations — in the query text; they become filters. If results miss, rephrase: raising effort only adds reading rounds.

Agentic search

effort	Gather rounds	First token	Reach for it when
default	1–2	3–4s	You want an answer now
high	3	7–8s	Comparing across papers
xhigh	4–5	9–13s	Survey-shaped questions

Rounds are a ceiling, not a floor — the loop converges early once the evidence is enough. Answer text starts within ~5s: a speculative prefetch removes the first evidence round, and the run is forced into the answer phase at force_answer_after.

Ask arXivregistered key+ /stream

POST/arxiv/agent/search

A question goes in; the service runs hybrid retrieval over the full-text corpus, reads the sections it needs, and returns an answer with real IDs — [arXiv:2512.15176]. Append /stream for NDJSON events.

CLIdeepxiv ask "…"

Parameters

Parameter	Type		Description
query	string	REQUIRED	1–2000 characters. Be specific; ask for numbers if you want numbers.
effort	default \| high \| xhigh	= default	Gather rounds ceiling: 1–2 / 3 / 4–5.
verbose	bool	= false	Include the tool-call trace (and `tool_call` / `tool_result` / `thinking` events when streaming).
top_k	integer	= 10	Speculative prefetch size, 1–30.
stream_answer	bool	= true	Streaming endpoint only. `false` emits one `answer` event instead of deltas.
max_rounds	integer	OPTIONAL	0–8. Overrides the effort preset. `0` answers straight off the prefetch — first token ≈1.3s.
force_answer_after	number	OPTIONAL	0.5–30s. Hard cutover into the answer phase.
max_answer_tokens	integer	= 4096	256–16384. Hitting it sets `answer_truncated`.
language	string	OPTIONAL	Answer language. Defaults to the query’s language.

Example

{
  "answer": "DEER reports a 5.54× speedup on HumanEval ... [arXiv:2512.15176]",
  "sources": [ { "arxiv_id": "2512.15176", "title": "...", "url": "https://arxiv.org/abs/2512.15176" } ],
  "stats": { "rounds": 1, "elapsed_s": 6.2, "answer_truncated": false },
  "trace": [ ... ],                       // only when verbose=true
  "quota": { "tier": "free", "used": 3, "limit": 30 }
}

Registered keys only. The token the SDK auto-registers returns 403. Every registered account gets 30 agentic calls a day, free, shared across arXiv, web and talent.

Citations are real. It never invents an ID and says “no relevant papers” rather than fabricate. sources is the retrieval set, not the citation list — filter to IDs that appear in answer. answer_truncated means incomplete — say so downstream.

Reading

One paper, seven views, all GET /arxiv/?type=…&arxiv_id=…. Each layer is a separate call; judge with brief, locate with head, read one section, and only then raw. Real numbers for 2409.05591: brief ~300 tokens, head ~1.7k, section “2. Method” 5,919, raw 23,311.

brief

GET/arxiv/?type=brief&arxiv_id={arxiv_id}

Title, TLDR, keywords, publication date, citation count, GitHub URL. ~300 tokens — enough to judge whether the paper is worth reading.

CLIdeepxiv paper 2409.05591 --brief

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

Response

arxiv_idarXiv ID
titlePaper title
tldrGenerated one-paragraph summary
keywordsKeyword list
publish_atPublication date
citationsCitation count
github_urlCode repository, when found
src_urlPDF link

head

GET/arxiv/?type=head&arxiv_id={arxiv_id}

Structured metadata plus the section map: every section name with its own TLDR and token count, so an agent knows where the answer lives before paying for it. Total token_count included.

CLIdeepxiv paper 2409.05591 --head

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

Response

title / abstract / authorsAuthors carry name, orgs and optional misc (email, orcid)
sections[{ name, idx, tldr, token_count }]
token_countTokens in the full paper
venue / journal_name / citations / categories / publish_at / keywords / github_urlAs available

section

GET/arxiv/?type=section&arxiv_id={arxiv_id}§ion={section}

One section, in full, as clean markdown. Use the names from head.

CLIdeepxiv paper 2409.05591 --section "2. Method"

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
section	string	REQUIRED	Section name exactly as listed by `head`, e.g. `1. Introduction`.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

preview

GET/arxiv/?type=preview&arxiv_id={arxiv_id}

The first N characters of the paper. Returns is_truncated and total_characters.

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
characters	integer	= 10000	100–100000.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

raw

GET/arxiv/?type=raw&arxiv_id={arxiv_id}

The entire paper as markdown. Reach for it last.

CLIdeepxiv paper 2409.05591 --raw

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

json

GET/arxiv/?type=json&arxiv_id={arxiv_id}

The complete structured document: metadata and every section as JSON.

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

markdown (HTML view)

GET/arxiv/?type=markdown&arxiv_id={arxiv_id}

A rendered HTML page for reading in a browser. Open 2409.05591.

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

Search & retrieve

GET/arxiv/?type=retrieve&query={query}

Hybrid retrieval (dense + sparse over metadata, section and RoC indexes, optional fine rerank) over arXiv, bioRxiv and medRxiv. Returns a ranked list, not an answer — for a cited answer use Ask arXiv. Queries transformer, attention mechanism, large language model are free.

CLIdeepxiv search "agentic memory" --venue NeurIPS --venue-year 2025

Parameters

Parameter	Type		Description
query	string	REQUIRED	Up to 500 characters.
source	arxiv \| biorxiv \| medrxiv	= arxiv	ID field in results follows the source: `arxiv_id` / `biorxiv_id` / `medrxiv_id`.
top_k	integer	= 10	1–100.
offset	integer	= 0	Pagination, 0–10000.
authors	string[]	OPTIONAL	Filter and ranking signal. Repeat the param per value.
orgs	string[]	OPTIONAL	Organisation filter and ranking signal.
categories	string[]	OPTIONAL	e.g. `cs.AI`, `cs.CL`. Pure filter.
venue	string[]	OPTIONAL	NeurIPS, ICLR, CVPR… aliases resolve (NeurIPS ↔ NIPS). `venues` is an alias.
venue_year	integer	OPTIONAL	1900–2100, parsed from the venue string.
min_citation	integer	OPTIONAL	Citation floor. Pure filter.
date_search_type	exact \| after \| before \| between	OPTIONAL	Pair with `date_str`.
date_str	string[]	OPTIONAL	`YYYY` / `YYYY-MM` / `YYYY-MM-DD`; for `between` pass start and end.
search_funcs	string[]	OPTIONAL	Index types, default `metadata, section, roc`.
use_fine_rerank	bool	= true	Fine rerank after recall.
return_contents	bool	= false	Include matched section contents.
return_roc	bool	= false	Include the RoC list.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

Example

{
  "status": "success", "total_count": 3,
  "result": [ { "arxiv_id": "2506.18871", "score": 0.9475, "title": "...", "tldr": "...", "abstract": "...",
                "authors": [{ "name": "...", "orgs": ["..."] }], "url": "https://arxiv.org/abs/2506.18871",
                "date": "2025-06-23T17:38:54Z", "citation_count": 217, "categories": ["cs.CV"] } ]
}

Filters combine with AND. A narrow date window on a high citation floor can legitimately return 0 — loosen one. Legacy size, search_mode, bm25_weight, vector_weight, date_from, date_to are gone.

Trending signal

GET/arxiv/trending_signal?arxiv_id={arxiv_id}&token={token}

Social engagement for one paper: tweets, likes, views, replies, first and last seen. 404 means no engagement yet — normal.

CLIdeepxiv trending · deepxiv paper 2409.05591 --popularity

Parameters

Parameter	Type		Description
arxiv_id	string	REQUIRED	arXiv paper ID. `2409.05591` and `2504.21776` are free.
token	string	REQUIRED	Required for this endpoint.

Example

{ "arxiv_id": "2409.05591", "total_tweets": 150, "total_likes": 3200, "total_views": 25000, "total_replies": 45,
  "first_seen_date": "2024-09-05T10:30:00", "last_seen_date": "2024-09-10T14:20:00" }

PMC, bioRxiv, medRxiv

PubMed Central articles live under /pmc/ with the same type convention. bioRxiv and medRxiv are reached through retrieve with source=biorxiv|medrxiv. Open-access only.

PMC · head

GET/pmc/?type=head&pmc_id={pmc_id}

PubMed Central metadata: title, DOI, abstract, authors, categories, date. PMC544940 and PMC514704 are free.

Parameters

Parameter	Type		Description
pmc_id	string	REQUIRED	PMC ID.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.

PMC · json

GET/pmc/?type=json&pmc_id={pmc_id}

The complete structured article.

Parameters

Parameter	Type		Description
pmc_id	string	REQUIRED	PMC ID.
token	string	OPTIONAL	API token, or send `Authorization: Bearer`. Not required for free papers.