Agentic AI & RAG¶
This chapter describes the theoretical foundations and practical implementations of the different agent architectures in LLARS.
Overview¶
LLARS implements four core paradigms for the interaction of LLMs with external knowledge and tools:
┌─────────────────────────────────────────────────────────────────────────────┐
│ Agentic AI Paradigms │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ RAG │ │ ACT │ │ ReACT │ │ ReflAct │ │
│ │ │ │ │ │ │ │ │ │
│ │ Retrieval │ │ Action │ │ Reasoning │ │ Reflection │ │
│ │ Augmented │ │ Only │ │ + │ │ + │ │
│ │ Generation │ │ │ │ Acting │ │ Acting │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │ │ │ │ │
│ ▼ ▼ ▼ ▼ │
│ Single-turn Multi-turn Multi-turn Multi-turn │
│ Retrieval Iteration + Reasoning + Reflection │
│ │
└─────────────────────────────────────────────────────────────────────────────┘
Chapters¶
| Chapter | Description | Status in LLARS |
|---|---|---|
| RAG | Retrieval Augmented Generation with hybrid search | ✅ Production |
| ACT | Action-only agents without explicit reasoning | ✅ Production |
| ReACT | Reasoning + acting in an iterative loop | ✅ Production |
| ReflAct | Reflection-based reasoning with state grounding | ✅ Production |
Paradigm Comparison¶
| Aspect | RAG | ACT | ReACT | ReflAct |
|---|---|---|---|---|
| Reasoning | Implicit | None | Explicit (THOUGHT) | Explicit (REFLECTION) |
| Iterations | 1 | 1-10 | 1-10 | 1-10 |
| Interpretability | Medium | Low | High | Very high |
| Speed | Fast | Fast | Medium | Medium |
| Complex Questions | Limited | Limited | Good | Very good |
| State Awareness | No | No | Partial | Full |
Which Paradigm When?¶
┌─────────────────────────────────────────────────────────────────┐
│ Decision Tree │
└─────────────────────────────────────────────────────────────────┘
Is the question a simple fact lookup?
│
├── YES → RAG (single turn, fastest answer)
│
└── NO → Does it require multi-hop reasoning?
│
├── NO → ACT (fast, no explanation needed)
│
└── YES → Is transparency important?
│
├── NO → ReACT (good reasoning)
│
└── YES → ReflAct (best interpretability)
Architecture in LLARS¶
┌─────────────────────────────────────────────────────────────────────────────┐
│ User Query │
└─────────────────────────────────────────┬───────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────────────────┐
│ AgentChatService │
│ (agent_chat_service.py) │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ agent_mode = chatbot.prompt_settings.agent_mode │
│ │
│ ┌──────────────────────────────────────────────────────────────────────┐ │
│ │ switch(agent_mode): │ │
│ │ case 'standard': → _chat_standard() → RAG + direct LLM │ │
│ │ case 'act': → _chat_act() → Action loop │ │
│ │ case 'react': → _chat_react() → Thought-action loop │ │
│ │ case 'reflact': → _chat_reflact() → Reflection-action loop │ │
│ └──────────────────────────────────────────────────────────────────────┘ │
│ │ │
└──────────────────────────────────────────┼──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────────────────┐
│ Tool Execution │
├─────────────────────────────────────────────────────────────────────────────┤
│ rag_search() → ChromaDB + hybrid search │
│ lexical_search() → FTS5 BM25 index │
│ web_search() → Tavily API (optional) │
│ respond() → Final answer │
└─────────────────────────────────────────────────────────────────────────────┘
Configuration Per Chatbot¶
Each chatbot can be configured individually:
# Database: ChatbotPromptSettings
agent_mode: Enum['standard', 'act', 'react', 'reflact']
task_type: Enum['lookup', 'multihop']
agent_max_iterations: int # Default: 5
tools_enabled: List[str] # ['rag_search', 'lexical_search', 'respond']
web_search_enabled: bool # Optional Tavily integration
Further Reading¶
- Lewis et al. (2020): RAG - Retrieval-Augmented Generation
- Yao et al. (2022): ReAct - Synergizing Reasoning and Acting
- Kim et al. (2025): ReflAct - World-Grounded Decision Making in LLM Agents via Goal-State Reflection