logosai-ontology 2.0.2

Knowledge-driven multi-agent orchestration with LLM-powered query analysis and intelligent agent selection

LogosAI Ontology System

Knowledge-driven multi-agent orchestration with LLM-powered query analysis and intelligent agent selection.

The Ontology System is the brain of the LogosAI platform. It analyzes user queries semantically, selects the optimal agent(s), designs execution workflows, and integrates results from multiple agents.

---

How It Works

User Query
    |
    v
Query Analysis (LLM) ──> Intent, entities, complexity
    |
    v
Agent Selection (Hybrid: Knowledge Graph + LLM)
    |
    v
Workflow Design ──> single | sequential | parallel | hybrid
    |
    v
Execution Engine ──> Agent calls with data piping
    |
    v
Result Integration ──> Unified response

Key Features

LLM-Based Agent Selection

• No hardcoded keyword matching -- all agent selection is semantic via LLM
• Agents are evaluated equally based on their metadata (description, capabilities, tags)
• When no suitable agent exists, the system provides constructive feedback instead of a wrong answer

Hybrid Agent Selection (v2.0)

Combines Knowledge Graph pattern learning with LLM reasoning:

Phase	Method	Purpose
1	Knowledge Graph	Entity extraction, pattern matching, time-decayed success rates
2	LLM Decision	Semantic analysis using graph insights + agent metadata
3	Feedback Loop	EMA success tracking, pattern generalization

Workflow Orchestration

Automatically determines the optimal execution strategy:

Strategy	When to Use	Example
single_agent	One agent can handle it	"What's the weather?"
parallel	Independent subtasks	"Search restaurants AND tourist spots"
sequential	Results feed forward	"Get price -> Convert currency"
hybrid	Mix of both	"(Weather || Exchange rate) -> Calculate expenses"

Agent Sync Service

Automatically synchronizes agent metadata from the ACP runtime server:

• Full sync on system startup
• File watcher detects new/changed agents every 5 seconds
• Updates Knowledge Graph, Agent Registry, and metadata in real-time

Project Structure

ontology/
├── core/                          # Core processing modules
│   ├── unified_query_processor.py # LLM-based unified query processing
│   ├── hybrid_agent_selector.py   # Knowledge Graph + LLM agent selection
│   ├── agent_sync_service.py      # Agent metadata synchronization
│   ├── llm_manager.py             # LLM client management
│   ├── llm_config_loader.py       # LLM configuration loading
│   ├── context_manager.py         # Query context management
│   ├── models.py                  # Data models
│   └── interfaces.py              # Abstract interfaces
│
├── engines/                       # Processing engines
│   ├── workflow_designer.py       # Dynamic workflow generation
│   ├── execution_engine.py        # Agent execution with data piping
│   ├── knowledge_graph.py         # Knowledge graph operations
│   ├── semantic_query_manager.py  # Semantic query handling
│   └── graph/                     # Graph engine and visualization
│
├── orchestrator/                  # Workflow orchestration
│   ├── query_planner.py           # LLM-powered execution planning
│   ├── execution_engine.py        # Multi-stage execution
│   ├── workflow_orchestrator.py   # Top-level orchestration
│   ├── progress_streamer.py       # Real-time progress streaming
│   ├── agent_registry.py          # Agent registration and discovery
│   ├── result_aggregator.py       # Multi-agent result aggregation
│   └── models.py                  # Orchestration data models
│
├── system/                        # System-level modules
│   ├── ontology_system.py         # Main ontology system
│   ├── knowledge_graph_manager.py # Knowledge graph lifecycle
│   ├── reasoning_generator.py     # Reasoning and inference
│   ├── result_integration.py      # Result integration logic
│   ├── strategy_manager.py        # Execution strategy selection
│   └── metrics_manager.py         # Performance metrics
│
├── services/                      # Support services
│   ├── agent_detector.py          # Agent capability detection
│   └── visualization_response_formatter.py
│
├── processors/                    # Query processors
│   └── enhanced_ontology_query_processor.py
│
├── config/                        # Configuration
│   └── llm_config.yaml            # LLM provider settings
│
├── utils/                         # Utilities
│   └── performance_analyzer.py
│
└── examples/                      # Usage examples
    ├── basic_usage.py
    └── advanced_usage.py

Quick Start

Prerequisites

• Python 3.11+
• Google API key (for Gemini LLM) or OpenAI API key

Installation

pip install -r requirements.txt

# Set API keys
export GOOGLE_API_KEY="your-google-api-key"

Basic Usage

import asyncio
from ontology.core.unified_query_processor import UnifiedQueryProcessor

async def main():
    processor = UnifiedQueryProcessor()

    available_agents = ['weather_agent', 'calculator_agent', 'search_agent']

    result = await processor.process_unified_query(
        query="What's the weather in Seoul?",
        available_agents=available_agents
    )

    print(f"Selected agent: {result['agent_mappings']}")
    print(f"Strategy: {result['execution_plan']['strategy']}")

asyncio.run(main())

Hybrid Agent Selection

from ontology.core.hybrid_agent_selector import get_hybrid_selector

selector = get_hybrid_selector()

# Select agent with Knowledge Graph + LLM
agent, metadata = await selector.select_agent(
    query="Show me Samsung stock price",
    available_agents=["search_agent", "finance_agent", "analysis_agent"],
    agents_info={...}
)

print(f"Selected: {agent} (confidence: {metadata['confidence']:.0%})")

# Store feedback for learning
await selector.store_feedback(query, agent, success=True)

Workflow Orchestration

from ontology.orchestrator import WorkflowOrchestrator

orchestrator = WorkflowOrchestrator()

# Process a complex multi-step query
async for event in orchestrator.process(query="Convert 100 USD to KRW and EUR, then compare"):
    if event["type"] == "progress":
        print(f"Stage: {event['stage']}")
    elif event["type"] == "final_result":
        print(f"Result: {event['result']}")

Configuration

LLM Settings (config/llm_config.yaml)

default_provider: gemini
providers:
  gemini:
    model: gemini-2.5-flash-lite
    temperature: 0.3
    max_tokens: 4096
  openai:
    model: gpt-4
    temperature: 0.3

Agent Selection Settings

Parameter	Default	Description
Time decay half-life	30 days	How quickly old patterns lose influence
Min weight	0.1	Minimum weight for old patterns
EMA alpha	0.3	Exponential moving average smoothing factor

Architecture

Core Design Principles

1. No Hardcoded Matching -- All agent selection uses LLM semantic analysis
2. Hybrid Intelligence -- Knowledge Graph patterns + LLM reasoning
3. Dynamic Workflows -- Execution strategies determined at runtime
4. Continuous Learning -- Feedback loop improves selection over time
5. Equal Agent Evaluation -- No agent is a "fallback"; all are scored equally

Integration with LogosAI

The Ontology System integrates with:

• ACP Server (port 8888) -- Executes selected agents
• logos_api (logosai-api) -- FastAPI backend
• logos_web (logosai-web) -- Next.js frontend
• logosai (logosai-framework) -- Agent SDK & runtime
• Knowledge Graph -- Stores and retrieves learned patterns

Related Repositories

Repository	Description	URL
logosai-framework	Python SDK for building agents	github.com/maior/logosai-framework
logosai-api	FastAPI backend server	github.com/maior/logosai-api
logosai-web	Next.js frontend	github.com/maior/logosai-web

Technology Stack

Component	Technology
Language	Python 3.11+
LLM	Google Gemini, OpenAI GPT
Knowledge Graph	NetworkX, custom graph engine
Query Processing	Async/await, aiohttp
Configuration	YAML, Pydantic

License

MIT License