Ask the first author.
Primary sources for agents. When an agent hits a question in a field, it comes here for the original: the paper, the statute, the judgment. It reads as much as it needs, in layers, and answers with citations you can open.
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 would do — badly, expensively, and without a citation you can check.
Search the web, then read a document
- Query a web search API. Ten links back — a few abstract pages, the rest blogspam.
- Open three of them. Guess from the summary which one holds the number, or the clause.
- Fetch the PDF or the page, parse it, watch the table — or the article numbering — collapse.
- Dump the whole document into context; hope attention finds the one line.
- Repeat for every baseline, every statute, every precedent you need.
- Assemble by hand. Attribute from memory. Hope the ID was not invented.
Ask the corpus a question
- →It picks its own tools. Hybrid retrieval over the full text of one domain, then reads the sections — or the articles — it needs.
- →Citations resolve. [arXiv:2512.15176] and [刑法 第266条] map to real documents. It says “nothing relevant” rather than inventing one.
- →It streams. First token at 3.5s p50 — answer to stdout, sources and progress to stderr.
- →Same shape for every domain. One request body, one NDJSON event protocol, one quota. Add a domain, not a new integration.
One vertical at a time, dug to the bottom
Exa searches the whole web. 1stAuthor digs one field until the primary sources are all there — then moves on to the next one. Expanding to more professional domains.
Papersformerly DeepXivLive
arXiv full text, PMC, bioRxiv and medRxiv. Search, judge, then read one section.
LawBuilding
Statutes from 26 jurisdictions and Chinese court judgments, cross-linked article ↔ case.
WebLive
General agentic search over the open web. Answers with URLs from cached page bodies, marking which were read in full.
TalentLive
Researcher profiles: affiliation, research tags, h-index, citations, representative papers, and a written survey per person.
Read in layers, not all at once
An agent deciding whether a document is worth reading shouldn’t pay for the whole document. Each layer is a separate call, cheap enough to run across the full candidate set. Real numbers.
2409.05591Judging whether this paper is worth reading costs 300 tokens, not 23,311 — 78× less.
（2022）湘0902刑初12号A statute reads the same way: retrieve → brief → article → context → raw. Article 266 alone is ~250 characters.
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, an article number or a case number — it says nothing relevant rather than fabricate. Tell your agent to preserve them in what it reports back.
Sources ≠ citations
A ten-document retrieval often supports one citation. sources is the retrieval set — filter to IDs that appear in the answer, or your agent will present unrelated documents 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 Law, every response carries a coverage block: Chinese civil judgments are ~12% indexed today and not a random sample. Put that in the tool description too.
Primary sources, parsed, kept current
Not a crawl of abstract pages. Full text, sectioned, embedded, re-synced.
| Source | Scale | Depth | Freshness |
|---|---|---|---|
| arXiv | 3,166,878 papers | Full text, sectioned, hybrid-indexed | T+0 — same day as announcement |
| PubMed Central | ~7.5M articles | Structured parse, section access | Daily |
| bioRxiv / medRxiv | Preprint corpus | Section access, same reading verbs | Daily |
| Statutes · 26 jurisdictions | 33,103 laws · 1,537,422 articles | Original + zh/en translation + per-article extraction | Building |
| Chinese judgments | 4,510,354 → 17.7M | Structured brief + opinion / decision sections | Building |
Have a corpus? We’ll make it a domain.
The pipeline behind Papers and Law — parse, section, extract, embed, cross-link, expose as reading verbs — is not specific to papers or law. Two ways in:
Community
Contribute a public corpus and we index it as a shared domain, credited, with the dataset released on seed.ac.cn.
Private
Hand us private data — internal documents, filings, manuals, records — and get a private 1stAuthor: same layers, same citations, same MCP tools, your keys only.
Give your agent something to reason over
Free to start. No card. Papers today, Law next, one API key for all of it.