Retrieval & language models
Paper Lens
Search a collection of research papers and trace every answer back to the passage that supports it.
The link opens my GitHub profile until a project repository is added.
The problem
A folder full of papers is easy to collect and difficult to search. The proposed tool would help a reader locate an explanation across documents, while preserving enough context to check the source.
How it would work
Extract text with document and page identifiers, then split it at section and paragraph boundaries. Keep the original passage alongside each embedding so search results remain inspectable.
Compare a lexical search baseline with dense retrieval. Feed a small set of relevant passages to the answer model, and display those passages beside the response instead of hiding them behind a citation badge.
Include questions that the collection cannot answer. The interface should make missing evidence visible and let readers open the source before trusting a generated explanation.
What to evaluate
A proposed evaluation plan for this example:
- Build a small, manually reviewed set of questions and relevant passages.
- Measure retrieval recall at a fixed k before changing the answer prompt.
- Review whether each answer is supported by its cited passages, and record latency alongside quality.