← All work

Research Paper Assistant

Question answering over a library of computer-vision papers, with structure-aware chunking, hybrid retrieval, reranking and answers that cite their sources.

Role: Developer
Team: Solo
Status: Runs locally
Links: Code ↗

Research Paper Assistant answering a question about DETR — Asking a question across the paper library (here, what loss and matching strategy DETR uses) returns an answer grounded in retrieved passages.

Problem

Papers break naive chunking: a fixed window cuts through equations, tables and figure captions, and the retrieved fragment then answers nothing.
For research questions, a plausible answer is worthless without the passage it came from.

What I built

PDF ingestion that follows document structure (sections, captions, references), so each chunk is a coherent piece of the paper.
Hybrid retrieval combining embeddings with keyword search, which keeps exact terms like model and dataset names findable.
Reranking on chunk metadata before the context is assembled, and answers returned with citations back to paper and section, served through FastAPI.

Key decisions

Hybrid over embeddings-only retrieval: model names and metrics are the queries that matter most, and they're lexical.
No answer without a citation. If retrieval comes back empty, the system says so.

Stack

Python
FastAPI
Embeddings
Hybrid retrieval
ChromaDB
PDF parsing