Skip to main content

hackagent.attacks.techniques.indirect.rag.retrieval

Chunking and cosine retrieval over the vectors an embedder returns.

Kept apart from the attack so the numeric parts — chunk offsets, FAISS, cosine — stay testable without a model, and so the attack reads as the algorithm rather than the linear algebra.

chunk​

def chunk(text: str, size: int, overlap: int) -> list[str]

Split text into overlapping chunks.

most_similar​

def most_similar(query: Sequence[float],
                 candidates: Sequence[Sequence[float]]) -> int

Index of the candidate vector closest to query by cosine.

top_k​

def top_k(query: Sequence[float], corpus: Sequence[Sequence[float]],
          k: int) -> list[int]

Indices of the k corpus vectors closest to query, best first.