ReACT - Reasoning and Acting

Theory

Paper

Original paper

Yao, S., Zhao, J., Yu, D., et al. (2022) ReAct: Synergizing Reasoning and Acting in Language Models DOI: 10.48550/arXiv.2210.03629 ICLR 2023

Concept

ReAct interleaves reasoning and acting in a loop. The LLM alternates between THOUGHT (reasoning), ACTION (tool call), and OBSERVATION (result). This enables transparent, traceable reasoning for complex multi‑hop requests.

Architecture

flowchart LR
    query([Query]) --> service[AgentChatService]
    service --> react[ReAct Mode]
    react --> thought[THOUGHT]
    thought --> action[ACTION]
    action --> tool[[Tool Execute]]
    tool --> obs[/OBSERVATION/]
    obs --> decide{Enough info?}
    decide -->|No| thought
    decide -->|Yes| final([FINAL ANSWER])

    style query fill:#98d4bb,stroke:#6bbf9a,color:#000
    style service fill:#a8c5e2,stroke:#7ba3c9,color:#000
    style react fill:#D1BC8A,stroke:#b8a06a,color:#000
    style thought fill:#b0ca97,stroke:#8fb077,color:#000
    style action fill:#D1BC8A,stroke:#b8a06a,color:#000
    style tool fill:#b0ca97,stroke:#8fb077,color:#000
    style obs fill:#88c4c8,stroke:#5fa8ad,color:#000
    style decide fill:#9e9e9e,stroke:#757575,color:#fff
    style final fill:#e8c87a,stroke:#d4a84b,color:#000

ReAct loop: Query → THOUGHT (forward‑looking) → ACTION → Tool → OBSERVATION → (repeat or answer)

Core Concept

THOUGHT → ACTION → OBSERVATION → THOUGHT → ... → FINAL ANSWER

The THOUGHT step is forward‑looking:

• "What do I need to do next?"
• "Which information is missing?"
• "Which tool should I use?"

Advantages over ACT

Aspect	ReACT	ACT
Transparency	High (THOUGHT visible)	Low
Debugging	Easy	Hard
Multi‑hop reasoning	Good	Limited
Self‑correction	Possible	Hard
Token usage	Higher	Lower

Typical THOUGHT Patterns

THOUGHT: I need to search for the opening hours first.
THOUGHT: The search returned no results, I will try different keywords.
THOUGHT: I now have all necessary information and can answer.

---

Example

Query

"When is DG Agentur open and how can I reach them?"

ReAct Trace

[Iteration 1]
THOUGHT: The user asks for opening hours AND contact details.
         I will search for opening hours first.
ACTION: rag_search("opening hours DG Agentur")
OBSERVATION:
  [1] Contact: Mon-Fri 9-18, Sat by appointment
  [2] You can reach us at +49 911 97554990

[Iteration 2]
THOUGHT: I found opening hours. For contact details,
         I will search for more information.
ACTION: rag_search("contact phone email address")
OBSERVATION:
  [1] Legal notice: Phone +49 911 97554990, Email office@dg-agentur.de
  [2] Address: Bahnhofstraße 2, 90402 Nürnberg

[Iteration 3]
THOUGHT: I now have all information: opening hours and
         full contact details. I can answer.
FINAL ANSWER: DG Agentur is open Monday to Friday 9-18,
              Saturday by appointment. [1]

              Contact:
              - Phone: +49 911 97554990
              - Email: office@dg-agentur.de
              - Address: Bahnhofstraße 2, 90402 Nürnberg [2]

Response

DG Agentur is open Monday to Friday 9-18,
Saturday by appointment. [1]

Contact:
- Phone: +49 911 97554990
- Email: office@dg-agentur.de
- Address: Bahnhofstraße 2, 90402 Nürnberg [2]

Sources:
[1] Contact - DG Agentur
[2] Legal notice - DG Agentur

---

Implementation in LLARS

Status: Production

ReACT is fully implemented and in production use.

Architecture

flowchart TB
    subgraph main[chat_react Loop]
        direction TB
        start([Start]) --> thought[Generate THOUGHT]
        thought --> action[Generate ACTION]
        action --> parse[Parse Response]
        parse --> check{FINAL ANSWER?}
        check -->|No| exec[Execute Tool]
        exec --> history[Add to Steps]
        history --> thought
    end
    check -->|Yes| done([Finalize])

    style start fill:#98d4bb,stroke:#6bbf9a
    style thought fill:#b0ca97,stroke:#8fb077,color:#000
    style action fill:#D1BC8A,stroke:#b8a06a,color:#000
    style parse fill:#a8c5e2,stroke:#7ba3c9,color:#000
    style check fill:#9e9e9e,stroke:#757575,color:#fff
    style exec fill:#b0ca97,stroke:#8fb077,color:#000
    style history fill:#88c4c8,stroke:#5fa8ad,color:#000
    style done fill:#e8c87a,stroke:#d4a84b
    style main fill:#f5f5f5,stroke:#b0ca97

System Prompt

# DEFAULT_REACT_SYSTEM_PROMPT (db/models/chatbot.py)
"""
Du bist ein ReAct-Agent. Du denkst Schritt für Schritt und führst Aktionen aus.

## Zyklus (wiederhole bis fertig):
1. THOUGHT: Analysiere was du als nächstes tun musst
2. ACTION: Führe GENAU EINE Aktion aus
3. Warte auf OBSERVATION

## Verfügbare Aktionen (NUR diese!):
- rag_search("suchbegriff") - Semantische Dokumentensuche
- lexical_search("suchbegriff") - Keyword-Suche
- respond("antwort") - Finale Antwort (beendet Prozess)

## Format (EXAKT einhalten!):
THOUGHT: [deine Überlegung]
ACTION: rag_search("suchbegriff")

Wenn fertig:
THOUGHT: [deine Überlegung]
FINAL ANSWER: [vollständige Antwort mit Quellen]

## WICHTIG:
- IMMER erst THOUGHT, dann ACTION oder FINAL ANSWER
- Aktionen GENAU so schreiben: rag_search("text")
- KEINE anderen Aktionen erfinden!
- Wenn keine Treffer: Query reformulieren, Komposita zerlegen und Synonyme testen.
"""

Additionally: - chatbot.system_prompt is prefixed. - build_tool_availability_prompt() adds the enabled tools dynamically. - {PROJECT_URL} placeholders are replaced before use.

Files

File	Function
app/services/chatbot/agent_chat_service.py	Routing to ACT/ReAct/ReflAct
app/services/chatbot/agent_modes/mode_react.py	chat_react() loop + streaming
app/services/chatbot/agent_parsers.py	parse_react_response()
app/services/chatbot/agent_tools.py	Tool execution + confidence checks
app/db/models/chatbot.py	DEFAULT_REACT_SYSTEM_PROMPT + prompt settings

Code Snippet

# mode_react.py - chat_react()
for iteration in range(max_iterations):
    yield {"status": "iteration", "iteration": iteration + 1, "max": max_iterations}

    # Stream THOUGHT + ACTION
    response_text, thought, action, final_answer = yield from _stream_react_response(...)

    # Final answer
    if final_answer:
        yield {"status": "final_answer"}
        ...
        return

    # Execute tool
    result, sources = service._tool_executor.execute_tool(action_name, action_param, message, enabled_tools)
    yield {"status": "observation", "result_preview": result[:300], "iteration": iteration + 1}

Parsing

# agent_parsers.py - parse_react_response()
THOUGHT_PATTERN = r"THOUGHT:\s*(.+?)(?=ACTION:|FINAL ANSWER:|$)"
ACTION_PATTERN = r"ACTION:\s*(.+?)(?=OBSERVATION:|FINAL ANSWER:|$)"
FINAL_PATTERN = r"FINAL ANSWER:\s*(.+?)$"

Configuration

# ChatbotPromptSettings
agent_mode: str = "react"
task_type: str = "lookup" | "multihop"
agent_max_iterations: int = 5

# Multihop: max_iterations = min(agent_max_iterations + 2, 10)

tools_enabled: List[str] = ["rag_search", "lexical_search", "respond"]
web_search_enabled: bool = False
web_search_max_results: int = 5

react_system_prompt: str = "..."  # custom prompt (optional)

Adaptive Iteration (High Confidence)

If the search yields high confidence, ReAct exits early and generates a final answer immediately. Confidence is derived from source scores (check_high_confidence).

Fallback Search

If no ACTION was generated but sources are required, ReAct triggers an automatic search (e.g., rag_search) to avoid stalling.

---

Events (WebSocket)

# Streaming Events (excerpt)
yield {"status": "starting", "mode": "react"}
yield {"status": "iteration", "iteration": 1, "max": 7, "steps": [...]}
yield {"status": "thinking", "iteration": 1}
yield {"status": "thought_delta", "delta": "...", "iteration": 1}
yield {"status": "thought", "thought": "...", "iteration": 1}
yield {"status": "action_delta", "delta": "...", "iteration": 1}
yield {"status": "action", "action": "rag_search", "param": "...", "iteration": 1}
yield {"status": "observation_delta", "delta": "...", "iteration": 1}
yield {"status": "observation", "result_preview": "...", "iteration": 1}
yield {"status": "adaptive_iteration", "iteration": 1, "reason": "high_confidence"}
yield {"status": "adaptive_response", "reason": "high_confidence_results"}
yield {"status": "max_iterations_reached"}
yield {"status": "final_answer"}
yield {"delta": "..."}
yield {"done": True, "full_response": "...", "sources": [...]} 

Logs

[AgentChatService] ReAct adaptive iteration: high confidence on iteration 2

Comparison: ACT vs ReACT in LLARS

Aspect	ACT	ReACT
Method	chat_act()	chat_react()
Location	mode_act.py	mode_react.py
THOUGHT step	No	Yes (streaming)
Adaptive iteration	Yes	Yes
Typical iterations	1-3	2-5

---

See Also

• ReflAct - Reflection-Grounded Agent Reasoning
• RAG - Retrieval Augmented Generation
• Agentic AI Overview