LLM decomposes complex business questions into logic dependency graphs, identifying dependencies between sub-problems.

Not one-shot answers, but breaking complex problems into manageable sub-problems for incremental solving.

1)    Problem identification: LLM analyzes the question, identifies sub-problems to solve.

2)    Dependency analysis: analyzes dependencies between sub-problems, determines solving order.

3)    Topological sort: topologically sorts sub-problems to ensure correct solving order.

4)    Context passing: previous sub-problem's solution feeds the next, passing context information.

Execute multi-step reasoning with framework constraints, topological sort plus context passing.

Each step has clear inputs, outputs, and reasoning rules, ensuring controllable reasoning.

1)    Framework constraints: reasoning executes within predefined framework constraints.

2)    Step-by-step reasoning: execute reasoning in topological order, each step with clear I/O.

3)    Context passing: previous step's results feed the next, maintaining reasoning coherence.

4)    Intermediate result caching: intermediate results can be cached to avoid recomputation.

Each reasoning step traceable, logic.trace records human-readable reasoning process.

Reasoning is not a black box, but a transparent chain that is auditable, verifiable, explainable.

1)    Reasoning records: each step records input, output, reasoning rules, execution time.

2)    Human-readable: records displayed in human-readable format for easy understanding and auditing.

3)    Reasoning replay: can replay the complete reasoning process, checking correctness step by step.

4)    Reasoning verification: can verify reasoning results to ensure correctness.

Predefined reasoning frameworks constrain the reasoning process for controllability and reliability.

Frameworks define available reasoning rules, constraints, and verification methods.

1)    Reasoning rules: frameworks predefine available rules, LLM reasons within rule scope.

2)    Constraints: frameworks define constraints preventing reasoning from going off track.

3)    Verification methods: frameworks define result verification methods for correctness.

4)    Framework extension: custom reasoning frameworks for different business scenarios.