ava-langchain-prompt-decomposition

Prompt Decomposition Engine (PDE) primitives for the Narrative Intelligence Stack.

Decomposes complex prompts through Four Directions analysis (Medicine Wheel epistemology):

• EAST (Vision/Spiritual): What is being asked? → Intent extraction
• SOUTH (Analysis/Mental): What needs to be learned? → Dependency mapping
• WEST (Validation/Emotional): What needs reflection? → Ceremony checks
• NORTH (Action/Physical): What executes? → Ordered action stack

Installation

pnpm add ava-langchain-prompt-decomposition

Quick Start

import { decompose } from 'ava-langchain-prompt-decomposition';

const result = decompose(
  'Research the existing patterns. Build the new module. Test the integration.'
);

console.log(result.json);     // Full PDE JSON
console.log(result.markdown); // Human-readable breakdown

Core Modules

| Module | Description | |--------|-------------| | DirectionalDecomposer | Classifies prompt segments into EAST/SOUTH/WEST/NORTH | | IntentExtractor | Extracts primary + secondary intents with confidence and urgency | | DependencyMapper | Builds dependency graphs with cycle detection and execution ordering | | ActionStackBuilder | Produces final PDE output (JSON/Markdown) | | MedicineWheelBridge | Direction↔Quadrant mapping, ceremony detection | | V0OntologyBridge | Maps to @medicine-wheel/* V0 package vision |

API

decompose(prompt, options?)

Convenience pipeline that runs all modules in sequence.

const result = decompose('Build a knowledge graph with ceremony.', {
  extractImplicit: true,
  maxItems: 20,
  ceremonyThreshold: 0.3,
});

result.decomposition  // DecompositionResult
result.wheelEnriched  // WheelEnrichedAnalysis
result.json           // JSON string
result.markdown       // Markdown string

DirectionalDecomposer

const decomposer = new DirectionalDecomposer();
const analysis = decomposer.decompose('Research and build.');

analysis.directions      // { east: [...], south: [...], west: [...], north: [...] }
analysis.leadDirection   // Direction.NORTH
analysis.balance         // 0.0 - 1.0
decomposer.isBalanced(analysis)       // boolean
decomposer.getGuidance(analysis)      // string[]

IntentExtractor

const extractor = new IntentExtractor({ extractImplicit: true });
const result = extractor.extract('Create a module and test it.');

result.primary     // { action: 'create', confidence: 0.9, urgency: 'session' }
result.secondary   // ActionIntent[]
result.context     // { filesNeeded: [...] }

Related Packages

• ava-langchain-relational-intelligence — MedicineWheelFilter, ValueGate, SpiralTracker
• ava-langgraph-prompt-decomposition-engine — Graph-level orchestration (depends on this package)
• ava-langchain-narrative-tracing — Narrative tracing handlers