README.md

September 4, 2026 · View on GitHub

DeepXiv

The data layer agentic search is missing — full paper text, real citations, and the people behind them.

Live system · Full documentation · API docs · Status · arXiv · 中文


deepxiv ask — a question in, a cited answer streaming out

What it is

A CLI and Python SDK over a service that has already read the literature. Papers arrive parsed into sections rather than PDFs, retrieval runs over full bodies rather than abstracts, and an agentic endpoint answers questions with citations that resolve to real arXiv IDs and URLs.

Four things it does:

Command
Ask the literaturedeepxiv askA question in, an answer out, cited with real arXiv IDs
Ask the webdeepxiv ask --webSame, over Google plus cached page bodies
Read a paper in layersdeepxiv search / paperSearch, judge, then read only the section you need
Find the peopledeepxiv talentWho works on a topic, where, and what their record is

What problem it solves

An agent researching a topic has bad options. Search APIs return ten blue links and abstracts — enough to name a paper, never enough to answer "what speedup does it report on HumanEval". PDFs answer that, but cost 50k tokens each and arrive as a wall of text with no structure to navigate.

DeepXiv removes the tradeoff. Papers are pre-parsed, so an agent can spend 300 tokens on a TLDR to decide whether to spend 5k on the Methods section. Questions that need evidence go to the agentic endpoint, which reads source text and hands back an answer with citations you can check. And because knowing who does the work is half of research, the same interface searches scholars.

Install

pip install deepxiv-sdk

Beta: deepxiv talent isn't on PyPI yet. It ships in 1.1.0b1 from source while the scholar index is still being built out:

pip install git+https://github.com/DeepXiv/deepxiv_sdk.git

deepxiv auto-registers a token on first use. Agentic commands (ask, talent) need a registered key instead — get one at data.rag.ac.cn/register, then:

deepxiv config --token YOUR_REGISTERED_KEY

Every account gets 300 agentic calls/day free, on a pool separate from the general daily limit.

Usage

One investigation, start to finish. You've heard speculative decoding got much faster this year and want to know what's real.

1. Ask the literature. Start with the question, not a keyword. The service picks its own tools, reads paper bodies, and cites what it used.

deepxiv ask "what speedup does speculative decoding report on HumanEval in 2025"

DEER reports a 5.54× speedup on HumanEval (with Qwen3-30B-A3B as the target
model), compared to EAGLE-3's 2.41× on the same benchmark [arXiv:2512.15176].

📚 Sources (1 cited, 10 retrieved — use --all-sources for the rest):
  1. [2512.15176] DEER: Draft with Diffusion, Verify with Autoregressive Models

The answer goes to stdout and sources to stderr, so deepxiv ask "…" > answer.md captures just the answer. Add --effort high when a question spans several papers.

2. Read the paper it cited — in layers. Never load a whole paper to answer a question about one section.

deepxiv paper 2512.15176 --brief              # title, TLDR, keywords, citations — worth reading?
deepxiv paper 2512.15176 --head               # section list + where the tokens are
deepxiv paper 2512.15176 --section Experiment # read only that

Each step costs an order of magnitude more than the last, so you stop as soon as you have your answer. Take section names from --head — papers don't share a common outline. --preview gives ~10k chars; no flag at all gives the full markdown.

3. Widen it into a search. Once you know what you're looking for, filter for the rest.

deepxiv search "speculative decoding" --date-from 2025-01 --min-citations 20 --limit 10

Filters combine with AND — --authors, --orgs, --categories, --venue/--venue-year, dates, citation floors. Stack too many and you'll legitimately get zero results; loosen one.

4. Find the people behind it. A method is worth more when you know whose lab it comes from and what else they've built.

deepxiv talent search "researchers working on speculative decoding" --semantic --limit 5
deepxiv talent survey 257                    # full profile: bio, education, work, open source, metrics
deepxiv talent survey 257 --format markdown  # the generated report

Semantic mode takes a sentence; drop --semantic to match names and affiliations directly. IDs from search feed survey.

5. Step off arXiv when the question isn't academic. Licensing, pricing, who shipped what last week — same command, different backend.

deepxiv ask "which inference providers support speculative decoding today" --web
deepxiv ask "NeurIPS 2025 best paper" --web --search-type news

The web backend reads cached page bodies. Pages read in full are marked 📄, snippet-only ones 🔗 — weigh them accordingly.

In Python, the same pipeline is Reader:

from deepxiv_sdk import Reader

reader = Reader(token="YOUR_REGISTERED_KEY")   # Reader takes the token explicitly
answer = reader.agent_search("what speedup does DEER report on HumanEval")["answer"]
method = reader.section("2512.15176", "Method")
people = reader.talent_search("speculative decoding", semantic=True, limit=5)

Documentation

Also available: PubMed Central, bioRxiv/medRxiv, trending papers, and per-paper social metrics — see USAGE.md.

Citation

If DeepXiv is useful in your work, please cite the technical report:

@article{qian2026deepxiv,
  title   = {DeepXiv-SDK: An Agentic Data Interface for Scientific Literature},
  author  = {Qian, Hongjin and Xia, Ziyi and Liu, Ze and Chen, Jianlyu and
             Luo, Kun and Qin, Minghao and Li, Chaofan and Xiong, Lei and
             Lan, Junwei and Wang, Sen and Liang, Zhengyang and Shao, Yingxia and
             Lian, Defu and Liu, Zheng},
  journal = {arXiv preprint arXiv:2603.00084},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.00084}
}

License & support

MIT — see LICENSE.