# Prompt 2: Reasoning Modellar va Fikrlash Zanjiri — Chain-of-Thought, Budget Tokens va Gallyutsinatsiyani Cheklash

> **Holat:** 2026-yil sentabr. Sun'iy intellekt sohasidagi eng katta inqilob — **Reasoning (fikrlovchi)** modellarining (OpenAI o1/o3, Claude 3.7 Thinking, Gemini Flash Thinking) paydo bo'lishidir. Ilgari model javobni tokenma-token keyingi so'zni bashorat qilish orqali zudlik bilan qaytargan bo'lsa, endi u javob berishdan oldin yashirin fikrlash fazasida rejalashtirish, xatolarni tekshirish va gipotezalarni tahlil qilish imkoniyatiga ega.

## Mundarija

1. [Reasoning modellar qanday ishlaydi: Test-Time Compute](#qanday-ishlaydi)
2. [Reasoning Tokenlar va Byudjet (`budget_tokens`)](#budget-tokens)
3. [Eski vs Yangi Prompting: Nimalardan voz kechish kerak?](#eski-vs-yangi)
4. [Chain-of-Thought (CoT) va Tree-of-Thoughts amaliyoti](#cot-amaliyoti)
5. [Gallyutsinatsiyani cheklash strategiyalari](#gallyutsinatsiya)
6. [Python orqali Reasoning API'ni boshqarish](#python-api)
7. [Xulosa](#xulosa)
8. [Manbalar](#manbalar)

---

## 1. Reasoning modellar qanday ishlaydi: Test-Time Compute {#qanday-ishlaydi}

An'anaviy LLM'lar (masalan GPT-4o) faqat o'qitish (pre-training) paytida o'rgangan bilimlariga tayanadi. Reasoning modellari esa **Test-Time Compute** (so'rov paytida hisoblash quvvatini oshirish) printsipi asosida ishlaydi:

```text
Oddiy LLM:     So'rov ─────────────────────────────► Darhol Javob (Fast)
Reasoning LLM: So'rov ──► [Ichki Fikrlash / Reja / Tekshirish] ──► Pishiq Javob (Slow & Deep)
```

Model Reinforcement Learning (RL) orqali kod kompilyatsiyasi, matematika va mantiqiy xulosalarda o'z javoblarini tekshirishga o'rgatilgan. Agar uning birinchi gipotezasi xato bo'lsa, u fikrlash zanjiri davomida o'zini o'zi tuzatadi (self-correction).

---

## 2. Reasoning Tokenlar va Byudjet (`budget_tokens`) {#budget-tokens}

2026-yilgi API'larda fikrlash chuqurligi dasturchi tomonidan aniq tokenlar soni bilan cheklanadi:

- **Anthropic Claude:** `thinking: { type: "enabled", budget_tokens: 4096 }`
- **OpenAI:** `reasoning_effort: "low" | "medium" | "high"`

Yashirin fikrlash tokenlari uchun ham odatiy token narxida haq olinadi, lekin ular foydalanuvchiga to'liq ko'rsatilmasligi yoki qisqartirilgan xulosa shaklida taqdim etilishi mumkin.

---

## 3. Eski vs Yangi Prompting: Nimalardan voz kechish kerak? {#eski-vs-yangi}

Reasoning modellari uchun prompt yozish an'anaviy modellardan tubdan farq qiladi:

| Eski Prompting (GPT-4 davri) | Reasoning Prompting (2026) | Sabab |
|---|---|---|
| "Qadamma-qadam o'yla (Think step by step)" | **Bunday ko'rsatma yozilmaydi** | Model shundog'i ham ichki fikrlash fazasiga ega; bu so'zlar uning fikrini chalg'itadi. |
| "Sen 20 yillik tajribali mutaxassissan" | **Aniq cheklovlar va kiruvchi ma'lumot beriladi** | Persona berish mantiqiy qobiliyatni oshirmaydi. |
| O'ta uzun va batafsil ko'rsatmalar | **Muammoning chekka holatlari (edge cases) va talablar aniq yoziladi** | Model o'zi eng optimal yechim yo'lini qidirib topadi. |

### Yaxshi Reasoning Prompt Strukturasi:
```text
<problem>
Ushbu Swift funksiyasi ko'p oqimli muhitda (Swift Concurrency) ma'lumotlar poygasi (data race) keltirib chiqarmoqda.
</problem>

<code_snippet>
final class AccountManager {
  var balance: Double = 0
  func deposit(amount: Double) { balance += amount }
}
</code_snippet>

<constraints>
- Swift 6 Complete Concurrency rejimiga to'liq mos bo'lsin.
- Lock yoki semaphore ishlatilmasin (faqat actor yoki Mutex).
- public API'ning mavjud imzosini buzmang.
</constraints>

<success_criteria>
Kod compiler xatolarisiz yig'ilishi va o'zgarishlar sababi tushuntirilishi shart.
</success_criteria>
```

---

## 4. Chain-of-Thought (CoT) va Tree-of-Thoughts amaliyoti {#cot-amaliyoti}

Murakkab arxitektura yoki xavfsizlik auditida modelga bir nechta muqobil yechimlarni taqqoslash vazifasini topshirish eng yuqori natija beradi:

```text
Gipotezalar daraxti (Tree of Thoughts) uslubida tahlil qil:
1. Ushbu iOS arxitektura muammosiga 3 ta muqobil yechim taklif qil (Masalan: VIPER, TCA, Clean Swift).
2. Har bir variantning kesh samaradorligi, testlash qulayligi va jamoaviy miqyoslanishini bahola.
3. Eng zaif tomonlarini ko'rsatib, yakuniy g'olib variantni tanla.
```

---

## 5. Gallyutsinatsiyani cheklash strategiyalari {#gallyutsinatsiya}

Model o'zidan yolg'on fakt to'qib chiqarmasligi uchun 3 ta amaliy qoida:

1. **Javob manbasini qat'iy cheklash (Grounding):**
   `"Faqat berilgan <context> ichidagi faktlarga tayangan holda javob ber. Agar ma'lumot bo'lmasa, 'Ma'lumot topilmadi' deb ayt."`
2. **Keltirilgan iqtiboslarni talab qilish (Attribution):**
   `"Har bir da'voingiz yonida kontekstdagi aniq qatordan iqtibos keltiring."`
3. **Tekshiruvchi (Critic) agenti:**
   Birinchi model javob yaratadi, ikkinchi model esa yaratilgan javobni birlamchi manba bilan satrma-satr tekshirib chiqadi.

---

## 6. Python orqali Reasoning API'ni boshqarish {#python-api}

Quyidagi Python skripti reasoning byudjetini hisobga olgan holda API so'rovini amalga oshirish logikasini ko'rsatadi:

```python
import json

def calculate_reasoning_cost(input_tokens, reasoning_tokens, output_tokens):
    # 2026 standart narxlari ($/1M token):
    input_price = 3.00
    reasoning_price = 15.00
    output_price = 15.00

    total_cost = (
        (input_tokens / 1_000_000) * input_price +
        (reasoning_tokens / 1_000_000) * reasoning_price +
        (output_tokens / 1_000_000) * output_price
    )
    return round(total_cost, 4)

# Misol hisob-kitob:
cost = calculate_reasoning_cost(input_tokens=1500, reasoning_tokens=3200, output_tokens=800)
print(f"Jami so'rov narxi: ${cost}")
```

---

## 7. Xulosa {#xulosa}

- Reasoning modellari murakkab muammolarni echishda sifatni yangi bosqichga olib chiqdi.
- "Think step by step" kabi eski prompt hiylalari endi kerak emas — o'rniga muammo chegaralari (constraints) va tekshirish mezonlari berilishi lozim.
- Byudjet tokenlarini to'g'ri boshqarish yuqori sifat va arzon narx o'rtasidagi muvozanatni saqlashga xizmat qiladi.

---

## 8. Manbalar {#manbalar}

1. [OpenAI — Learning to Reason with LLMs](https://openai.com/research/learning-to-reason-with-llms)
2. [Anthropic — Claude Thinking Mode Documentation](https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking)
3. [DeepSeek-R1 Technical Report](https://github.com/deepseek-ai/DeepSeek-R1)