Deterministic Reasoning Graph (DRG): A new paradigm for information organization

The Deterministic Reasoning Graph (DRG) is a novel paradigm for organizing information. DRG aims to eliminate indeterminism by explicitly structuring reasoning processes, ensuring non-probabilistic precision.

What DRG is not

• A new LLM architecture
• A learning algorithm
• A universal generalization solution

DRG structures information by capturing reasoning and decision-making patterns of a domain in a graph. It serves as a foundation for applications such as reasoning retrieval, controlled generation, or decision-making (agents).

Key use cases

• CoRG (Chain of Reasoning Graph): A structured, domain-specific RAG.
• Discriminator: Verification of an LLM’s output using a logical graph.
• Dataset Generator: Pre-applied reasoning to ensure domain-specific consistency.
• Conversational and Decision-Making Agent: Automation driven by reasoning and decision trees.

Understanding CoRG

This article focuses on CoRG and conversational/decision-making agents. Explicit reasoning relies on classification that enforces a non-probabilistic framework.

CoRG utilizes Chain of Tasks (cf. Episode 13). TL;DR: it’s a sequence of inferences.

In CoRG, the first inference is a context classification by the LLM. Then, this classification (label) is used to query a graph database. The retrieved context is injected into dynamic prompt tuning for the LLM before the second inference, enabling it to generate the output. This principle ensures the result, as the graph traversal is a query*.

Agentic architecture

On the agentic architecture side, CoRG is reused, but what differentiates it from a RAG is that it retrieves the reasoning and decisions of an agent.

LangGraph is similar but less granular:

• Routing Type: Explicit conditional (if, match, switch) vs. semantic and contextual
• Logic: Decision tree, routing by conditions (values, booleans, etc.) vs. decision tree applied to a graph, meaning no algorithmic limitations

Agent reasoning with DRG

Each agent relies on a DRG structure that doesn’t just connect steps conditionally but encodes semantic relationships like “triggered after,” “inherits from,” “modulated by,” or “validates.” This enables the agent to reason based on workflow patterns, operations, adaptability, and execution conditions. It transforms a system’s architecture into a controllable, reusable, and unambiguous logical graph.

Core principle of DRG

The graph is designed with business teams or by scientific domain according to an ontology specific to each use case. This structure constrains the LLM to reason explicitly based on domain-specific logic.

*In reality, queries also include nodes, relationships, and clusters. This is simplified here for clarity. To guarantee the result, the LLM must be fine-tuned.