Tool Use + Function Calling: LLM's Doors to External World — From OpenAI Tools to MCP

Ders Haritası (10 Bölüm)#

1. Tool Use motivasyonu — LLM'in 'eylemsizlik' problemi
2. OpenAI function calling — Haziran 2023 lansman
3. JSON schema tool definitions — name, description, parameters
4. Tool selection logic — model nasıl karar veriyor
5. Anthropic MCP — open standard Kasım 2024
6. Llama-3 tool tokens — open-source tool use
7. ReAct pattern — Reasoning + Acting iteratif
8. Plan-and-Execute — first plan, then execute
9. Reflexion — self-critique loop
10. Türkçe agent — production deployment

2-4. OpenAI Function Calling#

2.1 Function calling basic#

from openai import OpenAI
client = OpenAI()

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Belirtilen şehir için hava durumunu döndürür",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string", "description": "Şehir adı (örn: İstanbul, Ankara)"},
                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
                },
                "required": ["city"],
            },
        },
    }
]

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "İstanbul'da hava nasıl?"}],
    tools=tools,
    tool_choice="auto",
)

# Response.tool_calls içerir:
# {'function': {'name': 'get_weather', 'arguments': '{"city": "İstanbul"}'}}

2.2 Tool call cycle#

[1] User query → LLM
[2] LLM decides: tool needed?
   Yes → output tool_call (JSON with name + arguments)
   No → output text response
[3] Application executes tool, gets result
[4] Result sent back to LLM as tool message
[5] LLM combines: final response to user

2.3 Multi-tool, parallel tool calls#

User: 'İstanbul ve Ankara'da hava nasıl?'
LLM: [get_weather('İstanbul'), get_weather('Ankara')]  # parallel

2.4 Tool selection math#

5-9. MCP + Agent Patterns#

5.1 Anthropic MCP (Model Context Protocol, Kasım 2024)#

{
  "name": "github_search",
  "description": "Search GitHub repositories",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": {"type": "string"},
      "language": {"type": "string"}
    }
  }
}

5.2 MCP ecosystem 2025#

5.3 Llama-3 tool use#

<|python_tag|>{"name": "get_weather", "parameters": {"city": "İstanbul"}}<|eom_id|>

5.4 ReAct pattern (Yao 2022)#

Thought: 'Need weather info for trip planning'
Action: get_weather('İstanbul')
Observation: '22°C, sunny'
Thought: 'Good weather. Now check Ankara.'
Action: get_weather('Ankara')
Observation: '15°C, cloudy'
Thought: 'Have data for both. Final response.'
Answer: 'İstanbul 22°C güneşli, Ankara 15°C bulutlu.'

5.5 Plan-and-Execute#

Plan:
  1. Get weather for İstanbul
  2. Get weather for Ankara
  3. Compare and summarize

Execute step 1...
Execute step 2...
Execute step 3...
Final answer.

5.6 Reflexion (Shinn 2023)#

Attempt 1: Output X
Critic: 'Output doesn't address user's question fully.'
Reflect: 'Need to include source information.'
Attempt 2: Output Y (improved)

5.7 Production agent: Türkçe assistant#

• search_events(city, date)
• get_weather(city)
• send_email(to, subject, body)

Modül 20 Envanteri (Tamamlandı)#