Semantic Retrieval Poisoning in Lightweight RAG Systems

Semantic Retrieval Poisoning in Lightweight RAG Systems explores whether malicious documents can manipulate retrieval results in lightweight Retrieval-Augmented Generation (RAG) pipelines built with TinyLlama, ChromaDB, and sentence-transformer embeddings.

The project evaluates how poisoned documents influence semantic retrieval rankings and whether retrieval compromise leads to downstream instruction hijacking. Experimental results showed that poisoned documents frequently achieved high retrieval rankings, but successful retrieval manipulation did not consistently produce malicious model behavior.

The study highlights an important distinction between retrieval compromise and full behavioral compromise in resource-constrained language models. Rather than claiming complete system takeover, the work focuses on practical limitations, partial robustness, and measurable security boundaries within lightweight RAG architectures.