What I built
The approach
A question-answering system over 73 AI research papers: hybrid search, reranking, and a dashboard that tracks answer quality and speed.
Read the source ↗Reported evaluation
Repository-reported results across 10 queries using an LLM judge on a 1–5 scale. This small sample is exploratory; raw per-query results and a larger human-reviewed set are still needed for independent validation.
Follow the workflow
A research question starts the retrieval workflow.
Inside the project
Saved screenshots from the repository. Select an image to inspect it in detail.
Engineering decisions
Search by keywords and meaning
Keyword search catches exact technical terms while embeddings find passages with similar meaning.
Rerank before answering
Retrieves 20 candidates, then a second model chooses the best three for the final answer.
Show quality over time
Every query, response time and judge score is logged, surfacing weak answers and corpus gaps.
Limitations & next steps
Commit the raw 10-query results, document the judge rubric and model settings alongside them, connect RAGAS to the main dashboard path, and expand to a human-reviewed test set. Treat the reported latency as run-specific, not a service guarantee.