@metaventionsai/cpb-core
v1.0.1
Published
Cognitive Precision Bridge - Unified AI orchestration with precision-aware routing through RLM, ACE, and DQ scoring
Downloads
8
Maintainers
metaventionsai.comReadme
Architecture
┌─────────────────────────────────────────────────────────────────────────────┐
│ COGNITIVE PRECISION BRIDGE (CPB) │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────────────────────────────────────────────────────────────┐ │
│ │ QUERY ANALYSIS │ │
│ │ │ │
│ │ Input ──→ [Complexity Signals] ──→ [Path Scoring] ──→ Decision │ │
│ │ │ │ │ │
│ │ ┌───────────┴───────────┐ ┌──────┴──────┐ │ │
│ │ │ • Token count │ │ Score paths │ │ │
│ │ │ • Code indicators │ │ Consider │ │ │
│ │ │ • Reasoning patterns │ │ alternatives│ │ │
│ │ │ • Consensus signals │ │ Explain │ │ │
│ │ │ • Domain complexity │ │ reasoning │ │ │
│ │ └───────────────────────┘ └─────────────┘ │ │
│ └─────────────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────────────┐ │
│ │ EXECUTION PATHS │ │
│ │ │ │
│ │ ┌──────────┬──────────┬──────────┬──────────┬──────────┐ │ │
│ │ │ DIRECT │ RLM │ ACE │ HYBRID │ CASCADE │ │ │
│ │ │ <0.2 │ 0.2-0.5 │ 0.5-0.7 │ >0.7+ │ >0.7 │ │ │
│ │ ├──────────┼──────────┼──────────┼──────────┼──────────┤ │ │
│ │ │ Simple │ Context │ Consensus│ Combined │ Full │ │ │
│ │ │ queries │ compress │ building │ RLM+ACE │ pipeline │ │ │
│ │ ├──────────┼──────────┼──────────┼──────────┼──────────┤ │ │
│ │ │ ~1s │ ~5s │ ~5s │ ~10s │ ~15s │ │ │
│ │ ├──────────┼──────────┼──────────┼──────────┼──────────┤ │ │
│ │ │ Sonnet │ Sonnet │ Opus │ Opus │ Opus │ │ │
│ │ └──────────┴──────────┴──────────┴──────────┴──────────┘ │ │
│ └─────────────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────────────┐ │
│ │ ACE 5-AGENT ENSEMBLE │ │
│ │ │ │
│ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │
│ │ │ 🔬 │ │ 🤔 │ │ 🔄 │ │ 🛠️ │ │ 🔭 │ │ │
│ │ │ Analyst │ │ Skeptic │ │Synthesizr│ │Pragmatist│ │ Visionary│ │ │
│ │ │ │ │ │ │ │ │ │ │ │ │ │
│ │ │ Evidence │ │ Risks │ │ Patterns │ │ Feasible │ │ Strategy │ │ │
│ │ │ Logic │ │ Failures │ │ Connect │ │ Practical│ │ Long-term│ │ │
│ │ └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ │ │
│ │ │ │ │ │ │ │ │
│ │ └────────────┴────────────┼────────────┴────────────┘ │ │
│ │ ▼ │ │
│ │ [CONSENSUS ENGINE] │ │
│ │ Agreement scoring + synthesis │ │
│ └─────────────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────────────────┐ │
│ │ DQ SCORING │ │
│ │ │ │
│ │ ┌───────────────┐ ┌───────────────┐ ┌───────────────┐ │ │
│ │ │ VALIDITY │ │ SPECIFICITY │ │ CORRECTNESS │ │ │
│ │ │ 40% │ │ 30% │ │ 30% │ │ │
│ │ │ │ │ │ │ │ │ │
│ │ │ Addresses │ │ Detailed │ │ Factually │ │ │
│ │ │ the query? │ │ actionable? │ │ grounded? │ │ │
│ │ └───────┬───────┘ └───────┬───────┘ └───────┬───────┘ │ │
│ │ │ │ │ │ │
│ │ └─────────────────┼─────────────────┘ │ │
│ │ ▼ │ │
│ │ [OVERALL DQ SCORE] │ │
│ │ 0.75 threshold (ELITE) │ │
│ └─────────────────────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────────────────┘Features
- 5 Execution Paths: Direct, RLM, ACE, Hybrid, Cascade
- 5-Agent ACE Ensemble: Analyst, Skeptic, Synthesizer, Pragmatist, Visionary
- Smart Routing: Auto-selects optimal path based on query complexity
- Provider Agnostic: Works with any LLM (OpenAI, Anthropic, Gemini, etc.)
- DQ Scoring: Validity + Specificity + Correctness quality measurement
- Multimodal Support: Text and image inputs
- Real-time Status: Progress callbacks for UI integration
- ELITE TIER: Maximum quality configuration by default
Installation
npm install @metaventionsai/cpb-core
# or
yarn add @metaventionsai/cpb-core
# or
pnpm add @metaventionsai/cpb-coreQuick Start
import { createCPB, type CPBProvider } from '@metaventionsai/cpb-core';
// 1. Define your LLM provider
const claudeProvider: CPBProvider = {
name: 'claude',
isConfigured: () => !!process.env.ANTHROPIC_API_KEY,
generate: async (prompt, options) => {
const response = await anthropic.messages.create({
model: options?.model || 'claude-sonnet-4-20250514',
messages: [{ role: 'user', content: prompt }],
max_tokens: options?.maxTokens || 4096
});
return response.content[0].text;
}
};
// 2. Create CPB instance (ELITE TIER by default)
const cpb = createCPB({
fast: claudeProvider, // Sonnet for simple queries
balanced: claudeProvider, // Opus for RLM/ACE paths
deep: claudeProvider // Opus for cascade path
});
// 3. Execute with auto-routing
const result = await cpb.execute({
query: 'Compare microservices vs monolith architecture',
context: systemDesignDoc
});
console.log(result.output);
console.log(`Path: ${result.path}`); // 'ace'
console.log(`DQ Score: ${result.dqScore.overall}%`); // 78
console.log(`Confidence: ${result.confidence}`); // 85ELITE TIER Configuration
Default configuration optimized for maximum quality:
| Setting | ELITE Value | Standard Value | Description |
|---------|-------------|----------------|-------------|
| Default Path | cascade | direct | Full pipeline by default |
| Context Threshold | 100,000 | 50,000 | Chars for RLM activation |
| Complexity Threshold | 0.35 | 0.5 | Lower = more consensus |
| DQ Threshold | 0.75 | 0.6 | Minimum acceptable quality |
| Fast Path Time | 8s | 5s | More time for quality |
| Standard Path Time | 45s | 30s | Extended reasoning |
| Hybrid Path Time | 90s | 60s | Full pipeline allowance |
| RLM Iterations | 25 | 10 | Deeper decomposition |
| ACE Rounds | 18 | 8 | More consensus rounds |
| ACE Agent Count | 5 | 3 | Full ensemble |
Use Standard Tier (Cost-Conscious)
import { createCPB, STANDARD_CPB_CONFIG } from '@metaventionsai/cpb-core';
const cpb = createCPB(providers, STANDARD_CPB_CONFIG);5-Agent ACE Ensemble
The Adaptive Consensus Engine uses 5 specialized agents:
| Agent | Emoji | Role | Prompt Focus | |-------|-------|------|--------------| | Analyst | 🔬 | Evidence evaluator | Data, evidence, logical consistency | | Skeptic | 🤔 | Challenge assumptions | Failure modes, risks, edge cases | | Synthesizer | 🔄 | Pattern finder | Connections, frameworks, integration | | Pragmatist | 🛠️ | Feasibility checker | Actionability, resources, constraints | | Visionary | 🔭 | Strategic thinker | Long-term, second-order effects |
Consensus Scoring
Agreement is calculated via keyword overlap between agent responses:
// High agreement (>0.7): Strong consensus
// Moderate (0.4-0.7): Some divergence
// Low (<0.4): Significant disagreement - may need human reviewExecution Paths
| Path | Complexity | Use Case | Speed | Quality | Model | |------|------------|----------|-------|---------|-------| | Direct | <0.2 | Simple queries, navigation | ~1s | Good | Sonnet | | RLM | 0.2-0.5 | Long context, document analysis | ~5s | Better | Sonnet | | ACE | 0.5-0.7 | Decisions, trade-offs, consensus | ~5s | High | Opus | | Hybrid | >0.7 | Complex + long context | ~10s | Higher | Opus | | Cascade | >0.7 | Critical decisions, research | ~15s | Highest | Opus |
Path Selection Logic
// Automatic routing based on:
interface PathSignals {
contextLength: number; // Characters in context
queryComplexity: number; // 0-1 complexity score
requiresConsensus: boolean; // Multi-perspective needed?
requiresReasoning: boolean; // Deep analysis needed?
timeBudgetMs: number; // Time constraint
qualityTarget: number; // DQ threshold
}DQ Score Breakdown
interface DQScore {
overall: number; // 0-100 weighted average
validity: number; // 40% - Does it address the query?
specificity: number; // 30% - Is it detailed/actionable?
correctness: number; // 30% - Is it factually grounded?
}Quality Tiers
| Tier | Score | Status | |------|-------|--------| | Excellent | ≥0.85 | 🌟 | | Good | ≥0.75 | ✅ | | Acceptable | ≥0.60 | ⚠️ | | Below Threshold | <0.60 | ❌ |
Status Callbacks
const result = await cpb.execute(request, (status) => {
console.log(`Phase: ${status.phase}`);
console.log(`Progress: ${status.progress}%`);
console.log(`Path: ${status.path}`);
console.log(`Engine: ${status.currentEngine}`);
console.log(`Message: ${status.message}`);
});Phases
analyzing- Determining optimal pathcompressing- RLM context compressionexploring- Parallel explorationconverging- ACE consensus buildingverifying- DQ verificationreconstructing- Final synthesiscomplete- Done
Multi-Provider Setup
import { createCPB } from '@metaventionsai/cpb-core';
const cpb = createCPB({
fast: geminiFlashProvider, // Fast queries → Gemini Flash
balanced: claudeSonnet, // Analysis → Claude Sonnet
deep: claudeOpus // Deep reasoning → Claude Opus
});Router Utilities
import {
extractPathSignals,
selectPath,
canUseDirectPath,
needsRLMPath,
wouldBenefitFromConsensus
} from '@metaventionsai/cpb-core';
// Analyze without executing
const signals = extractPathSignals(query, context);
const decision = selectPath(signals);
console.log(`Recommended: ${decision.path}`);
console.log(`Confidence: ${decision.confidence}`);
console.log(`Reasoning: ${decision.reasoning}`);
console.log(`Alternatives:`, decision.alternatives);
// Quick checks
if (canUseDirectPath(query)) {
// Skip CPB, use direct LLM call
}
if (needsRLMPath(query, longContext)) {
// Context compression required
}
if (wouldBenefitFromConsensus(query)) {
// Multi-agent consensus recommended
}Force Specific Path
const result = await cpb.execute({
query: 'Design a new API',
forcePath: 'ace' // Force consensus path
});Research Foundation
CPB is built on research from:
| Paper | Topic | Application | |-------|-------|-------------| | arXiv:2512.24601 | Recursive Language Model | Context externalization, compression | | arXiv:2511.15755 | DQ Scoring | Quality measurement framework | | arXiv:2508.17536 | Voting vs Debate | Consensus strategies |
Related Packages
- @metaventionsai/voice-nexus - Voice AI architecture
- ResearchGravity - Python CPB implementation
License
MIT © Dicoangelo