Deterministic RAG: Beyond Probabilistic Retrieval

Abstract The traditional approach to RAG relies on the high-dimensional proximity of vector embeddings. While effective for semantic discovery, it introduces probabilistic uncertainty unsuitable for enterprise-grade compliance. This paper introduces the Dextar Deterministic Framework (DDF).

The Three-Layer Architecture Our architecture separates the concerns of episodic data storage, semantic identity mapping, and logical assertion verification.

1. Resource Layer (Immutable Fact Store) Every ingested document is assigned a content-addressed identity. This creates an infinite audit trail where every token generated by the system is linked to a specific resource hash.

2. Semantic Layer (Entity Mapping) Instead of relying on LLM internal weights, we construct an explicit knowledge graph that maps relationships between entities across documents.

3. Assertion Layer (Logic & Trust) Conflict resolution is handled via Quantitative Bipolar Argumentation (QBAF), ensuring that contradictory data points are resolved with a verifiable truth score.

Conclusion Moving beyond probabilistic retrieval ensures that AI systems can be trusted with sensitive organizational data.