Compositional reasoning
User-taught semantic composition and multi-hop inference without depending on a single memorized sentence.
SCHX / FEATURE LAYER
Source map
- python/abria_server/compositional_reasoning.py
- cpp/src/abr.cpp
SCHX / FEATURE LAYER
Purpose
compositional_reasoning.py builds grounded multi-step reasoning over learned structures so the system can combine compatible relations into a new answer.
SCHX / FEATURE LAYER
Bounded search
The module exposes configurable composition budgets and expansion caps to prevent user graph growth from causing unbounded request-thread stalls.
- ABRIA_COMPOSITION_BUDGET_MS defaults to 220
- ABRIA_COMPOSITION_EXPANSION_CAP defaults to 2400
SCHX / FEATURE LAYER
User-taught patterns
Procedural patterns learned from user examples are advisory. The reasoner can consider them while still cross-checking constraints and stronger evidence.