scite API 与 SciSpace API 对比：Smart Citations 与引用式答案

Choose scite API for a product centered on Smart Citation classifications. Choose SciSpace API for a research workspace that discovers papers, reads evidence, extracts fields, and writes cited answers. scite Assistant also generates referenced answers, so this is not a comparison of citation counts against text generation.

The deciding question is what your application must deliver beyond an answer: publication-level citation evidence, or a complete research workflow with tables and reports.

What are you actually comparing?

scite classifies citation statements as supporting, contradicting, mentioning, or unclassified. Its tallies count statements; citingPublications counts distinct citing publications. Those measurements are not interchangeable.

SciSpace offers two APIs. The Search API takes a research question and returns papers with full metadata. The Task API takes a plain-language research task; the agent plans the work, picks the tools, searches, reads, analyses and writes, then returns links to the files it created plus any supporting text answer.

Think of a citation-inspection panel versus a research workspace. Both can display cited prose, but the second also needs discovery, document handling, structured evidence, and follow-up tasks.

scite citation intelligence around a publication compared with a SciSpace research workflow producing cited answers, tables, and reports.

scite API vs SciSpace API at a glance

This compares documented capabilities, not measured accuracy or search completeness. scite also documents patent, grant, and clinical-trial evidence endpoints, so it should not be characterized as citation counts alone.

Dimension	scite API	SciSpace API
Literature coverage	Publication search and citation evidence	One search across SciSpace's 280M+ paper index, Google Scholar, PubMed and arXiv; a task can also search trial registries, Google Patents, grant databases and 77 science databases
Initial research call	Assistant task identifier	Search API returns papers; a Task API run streams its plan and progress live
Completed answer	Assistant turns with references	Links to every file the agent created plus any supporting text answer
Citation distinction	Classified citation statements and tallies	Paper-grounded answers, not scite classifications
Data extraction	Structured-response table mode	Extraction tables with AI-filled columns from full text; filter, sort and combine
Progress	Polling with intermediate information	Live streaming of each plan step and the answer as it is written

The scite Assistant guide documents its table mode and polling workflow; the task reference describes completed turns and references.

How should you test the same five questions?

Evaluation status: not executed. No authenticated outputs were supplied. These are shared test prompts and review criteria, not fabricated answers or a benchmark leaderboard.

Question for both APIs	Check in scite output	Check in SciSpace output
How does protein structure affect function?	Explanation and reference support	Explanation and citation support
What limitations of AlphaFold2 appear in studies published through 2025?	Evidence for each limitation	Evidence for each limitation
What supports or challenges the claim that larger language models always produce more factual answers?	Conflicting findings and qualifications	Conflicting findings and qualifications
How do lithium-ion battery recycling studies compare on recovery yield, process conditions, and limitations?	Comparable fields and units	Comparable fields and paper table
Which studies from 2020 through 2025 compare microplastic removal methods in wastewater, and what efficiencies do they report?	Date scope, methods, and denominators	Date scope, methods, and denominators

Use identical prompts, fresh conversations, and the same evidence inputs. Append: "Cite supporting papers, distinguish uncertainty, report units, and mark unavailable information as not reported."

Save raw outputs and settings. Review claim support, missing fields, conflicting evidence, and source accessibility.

How do you call both APIs?

The scite example is compact Python; the SciSpace side is described in plain language. Replace prompt with each test question and the shared instructions.

scite API Python request

The scite key needs the assistant scope, which is not included by default on a self-service Pro key. The endpoint reference uses userInput; this example follows that spelling rather than the guide's user_input example.

Poll GET /api_partner/assistant/tasks/{task_id} until SUCCESS, handling failure states separately. A task ID is not a completed answer.

What you send to SciSpace

Send the same question to the Search API to get a list of papers, each with all its relevant metadata. Send it to the Task API with the shared instructions, for example "answer with cited papers and build a table of recovery yield, process conditions and limitations", to get back links to the files the agent created plus any supporting text answer.

While a task runs, the plan and progress stream live, so your interface can show each step. Follow-up questions continue in the same thread.

What does each API return?

These are response shapes, not captured research results. scite's completed task contains assistant turns; retain their references rather than flattening everything into text.

Response stage	scite Assistant	SciSpace
Submission	`{"id":"example_task"}`	Task starts; plan and progress stream live
Completion	`status: "SUCCESS"`	Links to every file the agent created
Answer	`result.turns[].content`	Supporting text answer, plus Markdown, Word or PowerPoint files
Sources	`result.turns[].references`	Cited papers in the answer and files; the Search API returns papers with full metadata
User interface	Referenced answer; citation evidence where integrated	Cited answer, extraction table and generated files where requested

For the battery question, ask the Task API for an extraction table with columns for recovery yield, process conditions and limitations. The agent fills the columns from full text, and you can then filter, sort and combine the table.

For document-based evidence, SciSpace answers questions across up to 10 PDFs at once with page-level citations and reads figures and tables. See the PDF data extraction guide.

What does 1,000 scite API calls cost?

As checked on October 21, 2026, scite does not publish a separate per-call API rate from which to calculate this. Its Pro documentation lists $50 monthly or $480 annually for evaluation access; production permissions and expanded scopes require confirmation. These figures are not a price for 1,000 Assistant answers.

For scite's free access, public Papers and Tallies endpoints require no key. That is not a free Assistant allowance. scite limits depend on the endpoint and account; implement its documented retry handling.

Ask scite for terms covering your exact endpoints and workload. Define whether "1,000 calls" means submissions, polling requests, or completed research tasks before estimating anything.

When should you choose scite?

Choose scite when the product's defining feature is inspecting how publications cite another paper. Its statement classifications and tallies are directly relevant to publication-centered evidence panels.

Keep scite if those fields are part of your application contract. A SciSpace answer describing agreement or disagreement is not a replacement for scite's classification data.

Confirm deployment permissions: scite documents separate licensing for commercial or research use of Search and redacted Search snippets on self-service Pro keys.

If your shortlist began with elicit vs scite, compare API outputs rather than consumer interfaces. The Elicit API comparison offers another integration-focused evaluation path.

How do you migrate?

Treat migration as an output-contract change, not a base-URL swap. Preserve the question and acceptance tests while replacing job tracking and response parsing.

Migration from scite tasks to SciSpace threads while preserving research questions and adapting answer and source parsing.

scite parameter or concept	SciSpace mapping	Caveat
`turns`, `userInput`	The plain-language question or task	Preserve the research instructions
`sessionId`	Follow-ups in the same thread	Identifiers are not portable
Task `id`	A Task API run with live streaming	Change job tracking and completion checks
`answerLength`, `citationStyle`, `numReferences`	Plain-language instructions	Not equivalent dedicated controls
`yearFrom`, `yearTo`	State the date range in the request	Not a dedicated control
`useStructuredResponse`	Ask for an extraction table	Returned as a file, not the same schema

The scite parameters above follow its Assistant request reference.

SciSpace searches its own 280M+ paper index plus Google Scholar, PubMed and arXiv. It does not query Semantic Scholar, Scopus, Web of Science or Lens, so check coverage for the sources your scite results depend on.

Normalize answer text, sources, and status while retaining raw responses. Evaluate the five shared questions when access is available, then test follow-ups in the same thread and live progress streaming before changing production traffic.

Where the SciSpace API fits

scite is strongest at the citation layer: classified citation statements, tallies and referenced answers. If your product also needs study screening, extraction from full text and finished report files, you build that workflow around it.

The SciSpace API puts the full SciSpace Agent behind one call. It searches, reads, analyses and writes, and delivers the result as ready-to-use files.

API	What you send	What you get back
Search API	A research question in plain language, such as "limitations of AlphaFold2 reported through 2025"	A list of papers, each with all its relevant metadata, from one search across SciSpace's 280M+ paper index, Google Scholar, PubMed and arXiv
Task API (agentic)	A research task, such as a systematic review with a PRISMA flow diagram, an extraction table of recovery yield and process conditions, or a cited literature review	Links to every file the agent created (Markdown, Word, PowerPoint and more) plus any supporting text answer; the plan and progress stream live, and follow-ups continue in the same thread

What this adds on top of scite

One integration for the whole workflow: describe the task in plain language and the agent picks the tools (search, screening, extraction, analysis, writing).
Systematic review in one task: research questions, screening of up to 3,000 papers, extraction, PRISMA flow diagram and report.
Deeper discovery: deep literature review with up to 5 search-and-reflect rounds, and citation chasing forwards and backwards from seed papers.
Finished files: Markdown, Word, PowerPoint, or a LaTeX report compiled to PDF with its BibTeX.
Live streaming: each plan step and the answer stream as they are written, and follow-ups continue in the same thread.

When to use which

scite API: your product's defining feature is how publications cite a paper, using statement classifications, tallies and citingPublications.
SciSpace API: users need discovery, extraction and a cited deliverable in one workflow.

What it doesn't do

No citation-statement classification: SciSpace answers cite papers, but it does not label citations as supporting or contradicting the way scite does.
No paywalled papers: it works from your own PDFs plus 50M+ open-access full texts.
Only the sources listed here: it does not query Semantic Scholar, Scopus, Web of Science, Lens or other scholarly databases not named above.

The SciSpace API is in early access; it isn't open yet. At launch it's open to anyone: register on SciSpace and start, with no sales call. Every account gets $5 in free API credits, enough to try both APIs on real questions. The API is separate from SciSpace consumer plans; no consumer plan includes API usage.

Frequently asked questions

Last checked

October 21, 2026. scite details were checked against official documentation. No authenticated side-by-side benchmark was performed.

Sources

注意转载授权 · 阅读原文

scite API vs SciSpace API: Smart Citations vs Cited Answers

摘要

正文 · 原文