How SciContext will work
From a research question to a passage you can inspect.
SciContext is being built to help your AI find real papers, confirm their identity, and retrieve focused passages. The workflow below describes the planned public MCP, including its optional compact provenance receipt.
- Research question
- Relevant papers
- Selected paper
- Exact passages
- Optional receipt
Start with a question or a paper.
I have a research question
search_papers + query_paper
- 1Describe the subject or evidence you need.
- 2Your AI finds relevant papers through SciContext.
- 3Review the candidate papers and select the most useful ones.
- 4Retrieve passages from the selected paper.
I already have a paper
resolve_paper + query_paper
- 1Provide a supported paper identifier.
- 2SciContext confirms which work it refers to.
- 3Ask for passages relevant to your question.
- 4Inspect the text; request
prov: truewhen you need the compact receipt.
More than a confident paragraph.
SciContext returns passages, not a generated scientific conclusion.
Paper identity comes from search or resolve. A ready result includes passages with the applicable license and obligations. Adding prov: true adds one compact receipt for evidence inspection. Leaving prov out is equivalent to prov: false.
A retained benchmark returned five passages, with the expected answer term in the top-ranked passage. It evaluated passage retrieval, not the optional provenance receipt.
The optional receipt carries a bounded PROV-JSON document for the source-to-context derivation. It reports complete, partial, or unavailable according to the retained source lineage, without exposing SciContext's private corpus lineage. The receipt appears only on a ready prov: true result; non-ready states never fabricate one.
How is scaled dot-product attention computed?
"…compute the dot products of the query with all keys, divide each by √dₖ, and apply a softmax function…"
Attention Is All You Need · arXiv:1706.03762- Paper title
- Attention Is All You Need
- Authors and year
- Ashish Vaswani et al. · 2017
- SciContext paper ID
- scictx:paper:…
- Effective license
- CC-BY-4.0
- Obligations
- Attribution · License notice · Change notice
- Profile
- scicontext.query_paper.provenance · v2
- Status
- complete, partial, or unavailable
- PROV entities
- Retained source and canonical context
- Derivation
- Canonicalization activity, usage, and generation
- Source evidence
- URL, version, observation time, and optional SHA-256
Designed to make the source, and the limits, visible.
Stable paper identity
Reduce ambiguity between similarly titled works, versions, and identifiers.
Optional source-audit receipt
Request prov: true for the source-level receipt. Missing or ambiguous retained lineage is reported as unavailable, never invented.
Retraction-aware results
Known retractions are handled using the latest successful integrity observations from trusted scholarly sources.
Honest coverage
Some papers remain discoverable even when explicit source rights do not permit passage or image delivery. SciContext reports that limit instead of filling the gap with generated evidence.
Factual source attribution.
SciContext names the scholarly metadata and integrity sources it uses as factual attribution, without implying partnership or endorsement.
SciContext excludes papers marked retracted by its latest successful OpenAlex or Crossref integrity observation. Detection follows the freshness and coverage of those sources.
SciContext retrieves paper context through MCP.Your agent handles reasoning and synthesis.You make the scientific judgment.
SciContext
- Finds and identifies papers
- Returns passages with rights notices and optional provenance receipts
- Reports coverage limitations
Your agent
- Relates evidence to the question
- Compares findings
- Helps prepare an explanation
You
- Inspect the sources
- Evaluate limitations
- Make the final judgment
Give your AI a path back to the paper.
Move from plausible-sounding answers to scientific context you can inspect.