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)