Ingest PDFs or pull papers from academic databases into a knowledge graph of papers, authors, topics, and citations, then chat with your corpus: answers come only from the graph, with citations, or the tool says it does not know.
GRAPHRETRIEVEANSWER
corpus-only, by construction
Each of these exists in the codebase. We name the mechanism so you can hold us to it.
The model is constrained to the retrieved graph context: if the answer is not in your papers, it says so instead of improvising. Every answer returns its citations and the exact subgraph it used.
Papers, authors, topics, and keywords as nodes; CITES edges between papers; topics linked when they co-occur. Indexed and queryable, not a bag of embeddings.
Search across Semantic Scholar’s 200M+ paper corpus and OpenAlex, follow citation chains, and score candidates with five explainable signals: field overlap, co-citation, citation count, recency, title similarity.
Research gaps, topic clusters, field timelines, and collaboration patterns computed from the graph. Structure you would not spot reading paper by paper.
PDF upload with live progress; metadata extracted automatically.
Scope uploads, discovery, and the graph view to a collection: one corpus per project.
Force-directed graph with type filters, search fly-to, node detail, and export.
Results from multiple academic databases are deduplicated before they enter your graph.
Where it standsBeta, self-hostable with Docker, with 92 backend tests behind it. There is no hosted deployment yet. The graph covers what you ingest; search reaches Semantic Scholar’s corpus, we do not index 200M papers ourselves.
AI phone receptionist that answers calls and books the appointment.
Open-source AI support that resolves customer questions and escalates the hard ones.
MCP gateway that gates every agent tool call.