An open-source agent for scientific literature

An open-source agent for working with scientific literature, in active development right now — taking on the unsolved problem of searching inside the full text of a paper to find the passages that answer a question, not just its title and abstract.

AI research tools read the abstract and stop there. We are building an open-source agent for the part that stays hard even when the full text is right in front of you: searching and screening the whole paper, not just its abstract.

AI tools like Elicit, Consensus, and SciSpace help researchers and students review the scientific literature. They work mostly over abstracts, and they are closed. This post is about the problem that stays hard even when the full text is available — searching and screening the full text of papers, not just their abstracts.

01

The gap is measurable

Current tools screen titles and abstracts well, with reported sensitivity near 99%. Full-text screening is far weaker — specificity drops to roughly 47% — and it remains underexplored.

The screening gap

Title-and-abstract screening: reported sensitivity ~99%. Full-text screening: specificity ~47%, and largely underexplored.

Abstracts omit methods, sample sizes, tables, and author caveats, so a tool that stops at the abstract answers questions it cannot actually support. The problem is independent of access: even with the full text in hand, finding the right passage and deciding whether a paper truly meets a criterion is not solved.

02

What the agent does

The agent targets this problem directly. It retrieves papers from open databases — OpenAlex and Europe PMC — parses the full text with GROBID, and searches within the text for the specific passages that answer a query or satisfy an inclusion criterion, returning each one with its exact location.

The engineering is focused squarely on full-text search and screening: locating evidence spread across a paper, handling methods and tables, and reducing false inclusions.

03

What it can and cannot do

We report what the agent can and cannot do today — the trade-off between recall and precision in screening, and the failure modes we have observed. The contribution is an open-source alternative aimed at the part of the pipeline that proprietary tools have not solved and that does not go away, full-text search and screening, with an honest account of how far it reaches.

04

Why it matters

This matters most for education. Students rely on tools that read less of a paper than they think. An open agent that searches the full text, shows where each answer comes from, and stays honest about its limits gives them something they can actually check.

  • The hard, unsolved problem is searching and screening the full text — it stays difficult even when the PDF is available.
  • Abstract-only tools answer questions they cannot support, because abstracts omit methods, sample sizes, tables, and author caveats.
  • The gap is measurable: reported sensitivity ~99% on title-and-abstract screening versus specificity ~47% on full text.
  • The agent retrieves from OpenAlex and Europe PMC, parses with GROBID, and returns each answering passage with its exact location.
  • It is open source, focused on the part proprietary tools leave unsolved, and honest about its recall–precision trade-offs and failure modes.