Verified Retrieval

VaultCrux retrieval combines keyword matching with semantic vector search to find the most relevant chunks in your encrypted document corpus. Every query runs BM25 sparse search for exact term matching alongside dense vector search using Nomic 768-dimensional embeddings for semantic similarity. The results are merged through Reciprocal Rank Fusion (RRF) to produce a single ranked list that benefits from both signals.

Every result carries full retrieval metadata: the chunk ID, relevance scores, and the artifact provenance chain back to the original ingested document. This metadata is not just diagnostic - it flows through to proof jobs, so a cryptographic receipt can attest not only that an answer was generated but which retrieval path produced the evidence.

A cross-encoder reranker refines the final ordering by re-scoring candidates against the original query using a full cross-attention pass rather than the dot-product approximation used during initial retrieval. Rerank depth is configurable per retrieval mode: light mode defaults to 10 candidates, verified and audit modes to 100. The reranker consistently improves precision at the top of the ranked list.

The per-artifact cap prevents any single source document from dominating the result set, which is critical for corpora where one long document would otherwise monopolise the top positions. Combined with the deduplication pipeline that runs during ingestion, this ensures retrieval results are diverse by source, not just by content.

Retrieval modes - light, verified, and audit - control not just rerank depth but the entire quality-cost tradeoff. Light mode is designed for fast iteration during development. Verified mode activates stronger checks for sensitive outputs. Audit mode provides the highest traceability for compliance and high-stakes reviews, with every intermediate ranking decision logged for forensic replay.