Inspiration
Scientific literature often contains conflicting or condition-dependent findings, but conventional RAG systems tend to generate one confident answer. EvidenceSplit was inspired by the need to present research disagreements clearly and transparently.
What it does
Users enter a scientific claim and upload research papers. EvidenceSplit also searches for relevant papers online, extracts evidence, and groups findings into supporting, contradicting, and qualifying categories before generating a balanced comparison with citations.
How we built it
We used FastAPI, React, PostgreSQL with pgvector, OpenAlex, Unpaywall, PDF parsing, hybrid retrieval, and structured LLM outputs. Uploaded documents and live research sources are processed through evidence extraction, stance classification, and citation-grounded synthesis.
Challenges we ran into
The main challenges were extracting reliable evidence from different PDF formats, finding accessible full-text papers, distinguishing contradiction from qualification, and preventing unsupported conclusions or incorrect citations.
Accomplishments that we're proud of
We built a domain-agnostic pipeline that combines private uploaded documents with live scholarly sources and presents multiple sides of the evidence instead of forcing a single conclusion.
What we learned
We learned that scientific RAG requires more than semantic search. Reliable results depend on source provenance, evidence-level extraction, careful stance classification, and preserving the conditions and limitations reported by each study.
What's next for EvidenceSplit
We plan to add stronger reranking, automated evaluation, confidence scoring, study-quality indicators, larger document support, collaborative reports, and exportable literature-comparison summaries.
Log in or sign up for Devpost to join the conversation.