Smart Agent & Server

A high-performance, RAG-orchestrated LLM agent and OpenAI-compatible server with deep MCP integration.

Overview

This project provides a robust orchestration layer that transforms standard LLMs into specialized autonomous agents. It handles multi-turn tool loops, long-term memory via RAG, and serves everything through a standard OpenAI-compatible API.

Key Components:

• SmartAgent: The core orchestrator that manages intent classification, RAG retrieval, and tool execution loops.
• SmartServer: A production-ready HTTP server that makes the agent accessible to any OpenAI-compatible client (Cline, Goose, etc.).
• Hybrid RAG: A multi-vector search engine combining semantic embeddings with BM25 lexical scoring.

Features

• ✅ Real Incremental Streaming: True per-token streaming for both text and tool-call deltas.
• ✅ OpenAI SSE Compliance: Fully compatible with official OpenAI SDKs and IDE plugins.
• ✅ Hybrid RAG Search: Combines Vector similarity (semantic) with BM25 (lexical) for pinpoint accuracy in technical domains like SAP/ABAP.
• ✅ Multi-Intent Classification: Automatically routes requests to chat (fast-path), action (tool-loop), or long-term memory (fact/state).
• ✅ Reasoning Mode: Optional transparent thought process (<reasoning> blocks) to explain agent strategy.
• ✅ Resilience: Built-in exponential backoff retries for LLM/Embeddings and auto-reconnect for MCP servers.
• ✅ Helper LLM Support: Offload summarization and translation tasks to cheaper/faster models.
• ✅ Startup Health Checks: Immediate diagnostics for all dependencies (Ollama, MCP, LLM).

Installation

npm install @mcp-abap-adt/llm-agent

Quick Start

1. Initialize Configuration: Run the agent once to generate the default smart-server.yaml template:

   npx llm-agent-beta

2. Configure: Edit smart-server.yaml and .env with your API keys and MCP endpoints.

3. Launch Server:

   npm run dev -- --config smart-server.yaml

Usage

Connecting IDE Clients

Point your favorite AI client to your local SmartServer:

• Base URL: http://localhost:4004/v1
• Model ID: smart-agent
• API Key: (any string, e.g., sk-none)

Intent Types

SmartAgent automatically classifies your input into:

• Fact: Technical rules or constraints (e.g., "ABAP Cloud forbids direct table access"). Stored in long-term facts RAG.
• State: Project context or temporary observations (e.g., "Kristina approves decisions", "Sky is blue"). Stored in state RAG.
• Chat: Trivial questions or math (e.g., "2+2"). Processed via fast-path without project context.
• Action: (Default) Engineering tasks requiring tools and deep analysis.

Advanced Configuration (YAML)

port: 4004
host: 0.0.0.0

# Hybrid RAG Settings
rag:
  type: ollama
  url: http://localhost:11434
  vectorWeight: 0.7
  keywordWeight: 0.3

# Multi-model Pipeline
pipeline:
  llm:
    main:
      provider: deepseek
      apiKey: ${DEEPSEEK_API_KEY}
      model: deepseek-chat
    classifier:
      provider: deepseek
      apiKey: ${DEEPSEEK_API_KEY}
      model: deepseek-chat
      temperature: 0.1
    helper:
      provider: deepseek
      apiKey: ${DEEPSEEK_API_KEY}
      model: deepseek-chat  # Used for fast summarization/translation
      temperature: 0.1

# Feature Toggles
agent:
  showReasoning: true     # Enable debug thought blocks
  maxIterations: 10
  historyAutoSummarizeLimit: 15

Development

# Build project
npm run build

# Run tests
npm run test:all

# Development with hot-reload
npm run dev

# Smart server production entrypoint
npm run start

# Legacy compatibility aliases
npm run start:smart
npm run dev:llm
npm run start:llm
npm run test
npm run test:llm

License

MIT