DeepXiv 1.0agentic search is live

Filling in the data layer
agentic search is missing

Agents can reason. What they lack is a substrate to reason over: full paper text, real citations, and a retrieval loop that doesn't hand back ten blue links. Ask a question, get an answer grounded in sources you can verify.

$pip install deepxiv-sdk
Get an API key → Sign in Read the docs
3.13MarXiv papers, full text
T+0synced the day they're announced
3.5sp50 to first answer token
30/dayagentic calls, free
deepxiv ask — a question in, a cited answer streaming out
The workflow it replaces

You know the old loop. It costs fifty thousand tokens and a lot of trust.

Web search was built to hand a human ten links. Give that to an agent and it does what you'd do — badly, expensively, and without a citation you can check.

Before

Search the web, then read a PDF

  1. Query a web search API. Ten links back — a few abstract pages, the rest blogspam.
  2. Open three of them. Guess from the abstract which one holds the number.
  3. Fetch the PDF, parse it, watch Table 4 collapse into loose digits.
  4. Dump the whole paper into context; hope attention finds the one line.
  5. Repeat for every baseline you need to compare against.
  6. Assemble by hand. Attribute from memory. Hope the ID wasn't invented.
≈ 50k tokens · minutes · citations you can't verify
With DeepXiv 1.0

Ask the corpus a question

deepxiv ask "what compression ratio does KV cache eviction report on LongBench"
  • It picks its own tools. Hybrid retrieval over 3.1M full-text papers, then reads the sections it needs — not the abstracts.
  • Citations resolve. [arXiv:2512.15176] maps to a real paper. It says "no relevant papers" rather than inventing one.
  • It streams. First token at 3.5s p50 — answer to stdout, sources and progress to stderr.
  • Same shape for the open web. Add --web for pricing, products and news, from cached page bodies.
1 call · 3.5s to first token · every claim traceable
Two backends, one shape

Neither one is a wrapper around a search box

A question goes in; the service picks its own tools, reads sources when it needs to, and streams back a cited answer. The arXiv side reads paper sections; the web side reads cached page bodies.

arXiv deepxiv ask

Qdrant hybrid retrieval over a local full-text corpus, plus the paper bodies themselves. Answers with real IDs — [arXiv:2512.15176].

Best for — methods, reported numbers, experimental results, comparing what two papers claim on the same benchmark.

web deepxiv ask --web

Google plus cached page bodies. Answers with links to real URLs, and marks which pages were read in full versus snippet-only.

Best for — current events, products, pricing, anything that was never going to be in a paper.

EffortGather roundsFirst token · arXivFirst token · webReach for it when
default1–23–4s5–9sYou want an answer now
high37–8s≈13sComparing across papers
xhigh4–59–13slongerSurvey-shaped questions

Rounds are a ceiling, not a floor — it converges early once the evidence is enough. Raising effort adds reading rounds but can't redirect first-round recall; if an answer misses, rephrase more specifically instead.

Progressive reading

Read a paper in layers, not all at once

An agent deciding whether a paper is worth reading shouldn't pay for the whole paper. Each layer is a separate call, cheap enough to run across a full candidate set. Real numbers, from 2409.05591.

--briefTitle, TLDR, keywords, citation count, GitHub URL
~300 tok
--headSection map with per-section token counts — where the answer lives
~1.7k tok
--section "2. Method"One section, in full, as clean markdown
5,919 tok
--rawThe entire paper
23,311 tok

Judging whether this paper is worth reading costs 300 tokens, not 23,311 — 78× less. A question about its method costs 5,919. Search, judge, then read.

Search that filters like a database

Author, organisation, category, venue (NeurIPS ↔ NIPS aliases resolve), venue year, date ranges, minimum citations, offset paging, opt-in fine reranking.

Trending, by social signal

What's actually being read right now — deepxiv trending for the rolling window, --popularity for a single paper's views, tweets and likes.

Beyond arXiv

PubMed Central, bioRxiv and medRxiv share the same reading verbs. Open-access only — agents work on unrestricted data instead of hitting paywalls.

# search → judge → read: three calls, not one 23k-token dump deepxiv search "agentic memory" --venue NeurIPS --venue-year 2025 --min-citations 50 deepxiv paper 2409.05591 --brief deepxiv paper 2409.05591 --section "2. Method" # filters combine with AND — stacking a narrow date window on a high # citation floor can legitimately return zero. Loosen one.
What you can trust

Three things to know before you wire this into an agent

An agent handed a bare ask(query) tool will use this API poorly. These three distinctions belong in your tool descriptions.

Citations are real

Never invents an arXiv ID or URL — it says no relevant papers rather than fabricate one. Tell your agent to preserve them in what it reports back.

Sources ≠ citations

A ten-paper retrieval often supports one citation. sources is the retrieval set — filter to IDs that appear in the answer, or your agent will present unrelated papers as evidence.

Truncation is flagged, not hidden

Hit max_answer_tokens and the API sets answer_truncated. Surface it, or an agent will summarise a cut-off answer as though it were whole.

On the web backend evidence has two strengths: pages marked read were parsed in full, the rest gave only a search snippet. It never fetches live — an uncached page is weaker evidence.

The corpus

Open-access literature, kept current

Not a crawl of abstract pages. Parsed full text, sectioned, embedded, re-synced daily.

SourceScaleDepthFreshness
arXiv3,127,076 papersFull text, sectioned, hybrid-indexedT+0 — same day as announcement
PubMed Central~7.5M articlesStructured parse, section accessDaily
bioRxiv / medRxivPreprint corpusSection access, same reading verbsDaily
Open webGoogle indexCached page bodies, not live fetchCache-first
Tokens & limits

Two pools, independent of each other

Agentic calls don't consume your general limit, and vice versa. Agentic search needs a registered key — the SDK's auto-registered token isn't eligible and returns 403.

KeyGeneral requests / dayAgentic calls / dayHow to get it
Auto-registered1,000— not eligibleAutomatic on first CLI use
Registered10,00030, freedata.rag.ac.cn/signup →
Lite / PremiumCustom500 / 10,000Email tommy[at]chien.io with your use case

Lost your key? Sign in — your keys are on your dashboard. Free test papers, no token needed — arXiv 2409.05591, 2504.21776; PMC PMC544940.

Get started

Three steps to a cited answer

Install

CLI, the Reader client, and a token that registers itself on first use.

pip install deepxiv-sdk

Register for agentic search

Google sign-in or phone verification. Under a minute, and lifts you to 10,000 requests plus 30 agentic calls a day.

deepxiv config --token YOUR_KEY

Ask

Be specific — "what speedup does X report on HumanEval" beats "speculative decoding" by a wide margin. Chinese works directly.

deepxiv ask "..." --verbose

Give your agent something to reason over

Free to start. 3.1M full-text papers, real citations, an answer in under four seconds.