luki-core-agent

This repository is archived. Active development continues in a private repository with expanded persona management, improved safety detection, and streaming architecture. This public version reflects the ReMeLife integration era and is no longer maintained.

AI agent framework: dialogue orchestration, tool routing, safety filtering, and LLM backend integration.

Privacy Notice

Proprietary components have been removed from git history for public release:

• Persona definitions and personality frameworks
• Domain-specific knowledge bases and prompt templates
• Internal test scenarios with API keys

The core framework (orchestration, safety chain, tool routing, context building) remains intact. The service requires these proprietary modules (prompts/, _context/, knowledge/) to fully function.

What It Does

• Orchestrates multi-turn conversations with configurable personas
• Routes tool calls to downstream modules (cognitive, engagement, reporting)
• Filters content through a safety chain (pattern detection, PII awareness)
• Manages LLM backends with fallback support (Together AI, OpenAI)
• Builds context from ELR memories, conversation history, and persona prompts
• Streams responses via SSE

Stack

• Framework: FastAPI
• LLM: Together AI (primary), OpenAI (fallback)
• Prompts: Jinja2 templates
• Safety: Configurable content filtering and pattern detection
• Deployment: Docker on Railway

Structure

luki_agent/
├── dev_api.py            # FastAPI server (streaming + non-streaming chat)
├── config.py             # Feature flags, model routing, service URLs
├── llm_backends.py       # LLM provider abstraction (Together AI, OpenAI)
├── safety_chain.py       # Content filtering and safety compliance
├── context_builder.py    # Context assembly (ELR + history + persona)
├── context_optimizer.py  # Context caching and optimization
├── module_client.py      # HTTP client for downstream modules
├── project_kb.py         # Project knowledge base retrieval
├── prompt_registry.py    # Prompt template management
├── schemas.py            # Pydantic response models
├── resilience.py         # Retry + circuit breaker for HTTP calls
├── tools/
│   ├── registry.py       # Tool registration and routing
│   └── web_search.py     # Tavily web search integration
├── memory/
│   ├── memory_client.py        # ChromaDB memory client
│   ├── memory_service_client.py # Memory service HTTP integration
│   ├── retriever.py            # Semantic search retrieval
│   └── session_store.py        # Ephemeral session memory
└── features/
    └── tiers.py          # Subscription tier logic

Prerequisites

This service requires running instances of:

• luki-memory-service — ELR memory storage (required for context building)
• Together AI API key — primary LLM provider
• Proprietary modules (prompts/, _context/) are required at runtime but not included in this public release

Setup

git clone git@github.com:ReMeLife/luki-core-agent.git
cd luki-core-agent
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

export TOGETHER_API_KEY=your_together_key
export MEMORY_SERVICE_URL=http://localhost:8002
uvicorn luki_agent.dev_api:app --reload --port 9000

Key Endpoints

Method	Path	Description
POST	/v1/chat	Non-streaming chat
POST	/v1/chat/stream	SSE streaming chat
GET	/health	Service health

Safety

The safety chain runs pattern-based detection on user input. Rules are defined in safety_chain.py and enforce content boundaries before sending context to external LLMs.

License

Apache License 2.0. Copyright 2025 Singularities Ltd / ReMeLife. See LICENSE.