The service currently runs on a single small machine: about 50 search requests/s and ~20 concurrent ask sessions across all users. Please keep concurrency modest (≤ 5 per client) and back off on 429 using Retry-After. Details →

Papers

Overview

arXiv full text, structured section by section with a TLDR for each; bioRxiv and medRxiv alongside it.

status
live
document id
arxiv_id e.g. 2409.05591
citation
[arXiv:2409.05591]
updated
2026-09

Common calls

Find papersPOST /v1/papers/search {"query": "speculative decoding", "filters": {"publish_at": {"gte": "2024"}, "categories": ["cs.CL"]}}
Read a paperGET /v1/papers/read/2409.05591
Just askPOST /v1/papers/ask {"query": "DEER 在 HumanEval 上加速多少", "effort": "low"}

The query matters more than the parameters. 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. Year, venue, category, author, institution and minimum citations belong in filters. If the results miss, rephrase — more effort only adds reading rounds.

The fine print
  • `section` (with level=full) must match sections[].name from level=detail exactly, numbering included (e.g. '1. Introduction').

Read

GET/v1/papers/read/{arxiv_id}
Three depths. detail is the default. Free sample: 2409.05591.

Levels

levelcreditsreturns
brief1Title, TLDR, keywords, date, citations, GitHub. ~300 tok
detail (default)2Metadata + section map (name / tldr / tokens per section) + total tokens
full5The whole paper as markdown

Extras

One side of the same document, in place of a level.

extracreditsreturns
trending1X/Twitter attention signal

Parameters

ParameterTypeDescription
arxiv_id (path) REQUIRED string Document id (arxiv_id).
level brief · detail · full= detail How deep to read.
extra trending Read one side instead of a level.
format json · md · html= json Response shape.
section string (level = full: read one section)
characters integer= 10000 (level = full: first N characters)

Example

{ "domain": "papers", "id": "2409.05591", "level": "detail", "view": "head", "data": { … } | "text": "…", "token_count": 1834, "meta": { … } }
The older GET /v1/papers/doc/{id}?view=… still works.

Read many

POST/v1/papers/read
Up to 50 ids, one level. Missing ids come back with error: "not_found" and are not charged.

Parameters

ParameterTypeDescription
ids REQUIRED string[] Document ids.
level brief · detail · full= detail Level for every id.
extra trending Read one side instead.

Example

{ "domain": "papers", "level": "detail", "docs": [ { "id": "…", "data": { … } }, { "id": "…", "error": "not_found" } ], "meta": { … } }

Facets

GET/v1/papers/facets?field={field}
Legal values for a filter field, so an agent never guesses; precomputed fields carry counts, declared enums come without. Free, no key.

Parameters

ParameterTypeDescription
field REQUIRED source · categories · venues Filter field with facets.
limit integer= 50 Max values.

Example

{ "domain": "papers", "field": "…", "values": [ { "value": "…", "count": 3120 } ], "meta": { … } }

Askregistered key+ /stream

POST/v1/papers/ask
Hybrid retrieval, reads sections, answers with real [arXiv:id] citations; /stream is NDJSON. effort low / medium / high = 50 / 100 / 200 credits. Append /stream for NDJSON; the first event is always billing. A failure before the first answer_delta is refunded.

Parameters

ParameterTypeDescription
query REQUIRED string 1–2000 characters. Be specific; ask for numbers if you want numbers.
effort low · medium · high= low Reading rounds, and the price: 50 / 100 / 200 credits. Your plan caps it.
verbose bool= false Include the tool-call trace (tool_call / tool_result / thinking events when streaming).
top_k integer= 10 Speculative prefetch size.
stream_answer bool= true Streaming only: false emits one answer event instead of deltas.
max_rounds integer 0–8, overrides the effort preset.
force_answer_after number Seconds; hard cutover into the answer phase.
max_answer_tokens integer= 4096 Hitting it sets answer_truncated.
language string Answer language; defaults to the query’s.

Example

{ "answer": "… [arXiv:2409.05591] …", "sources": [ … ], "stats": { "rounds": 1, "elapsed_s": 6.2, "answer_truncated": false }, "coverage"?: { … }, "meta": { "credits": { "charged": 50, … } } }
Citations are real — every id in answer exists in this domain (format [arXiv:2409.05591]). sources is the retrieval set, not the citation list. answer_truncated means incomplete.