An agentic research assistant for querying, comparing, and understanding CVPR papers with open-source models.
This project builds a local research assistant for CVPR literature using arXiv ingestion, FAISS retrieval, QLoRA-fine-tuned Qwen models, and a LangChain agent loop.
This is a research pipeline first, so Python and PyTorch define the project's feel across ingestion, vectorization, fine-tuning, and agentic querying. The result is closer to a technical research assistant than a basic document chatbot.
The system is designed as a locally hosted, agentic RAG assistant for CVPR paper comprehension.
Its pipeline covers collecting from arXiv, processing PDFs, chunking, embedding, FAISS indexing, fine-tuned Qwen model inference, and LangChain agent orchestration.
It also has a stronger evaluation layer than many portfolio projects by measuring both retrieval and generation quality.
Uses sentence-transformer embeddings with FAISS for efficient retrieval.
Applies QLoRA to fine-tune open-source Qwen models under realistic hardware constraints.
Wraps retrieval and answer generation in an agentic query flow rather than static retrieval-only answering.
Tracks hit rate, MRR, ROUGE, and semantic similarity to compare model iterations.
Documents multiple generations of the model rather than only a single final checkpoint.
Even the integration experiments expose useful engineering tradeoffs rather than being hidden away.
Multiple model iterations are compared quantitatively, which gives the project a much stronger evaluation backbone.
The engineering story is not only about success metrics but also about the tradeoffs that surfaced during integration.
Scrapes metadata and paper text from arXiv to create a dedicated CVPR retrieval corpus.
Uses sentence-transformer embeddings with FAISS for efficient context retrieval.
Wraps the RAG pipeline in an agentic workflow with retriever and generator evaluation.
Collect and process CVPR papers from arXiv into a local corpus.
Index document chunks in FAISS and fine-tune the model with QLoRA.
Answer user queries through a RAG plus agent loop and evaluate both retrieval and generation quality.
The project benefits from an iterative model story, where later versions clearly improve on earlier ones.
That makes the system feel more credible because it shows how the product matured over time.
The system is built around a clear product problem: dense research papers are difficult to search, compare, and synthesize quickly when the reading workflow is manual. By combining ingestion, chunking, retrieval, model fine-tuning, and agentic query handling, the product turns a static corpus into something that can support iterative understanding. The goal is not only search, but better comprehension of technical literature.
Its usefulness comes from the fact that retrieval is only the entry point. The system is built to help with comparison, explanation, and follow-up reasoning across a specialized research domain, which makes it more valuable than a document finder alone. The evaluation layer also adds credibility because it measures how well the assistant retrieves and explains information, not just whether it can produce fluent text.