Bài 19: LLM APIs — OpenAI, Anthropic Claude, Google Gemini

AI & LLM: Từ Cơ bản đến Nâng cao

Phần 5: Xây dựng AI Applications

xdev.asia

Tổng quan

Thay vì tự deploy model, bạn thường sẽ gọi LLM qua API. Bài này hướng dẫn thực tế với 3 provider lớn nhất: OpenAI, Anthropic và Google Gemini — covering streaming, vision, structured output và cost optimization.

---

1. So sánh Providers (2025)

Provider	Model	Context	Điểm mạnh
OpenAI	GPT-4o, o3	128K	Tool use, coding, ecosystem lớn
Anthropic	Claude Opus/Sonnet	200K	Long context, safety, reasoning
Google	Gemini 2.0	1M	Multimodal, free tier, long context
Meta (via API)	LLaMA 3.3	128K	Open weights, self-hostable
Mistral	Mistral Large	128K	EU data, affordable

---

2. OpenAI API

Setup

pip install openai
export OPENAI_API_KEY="sk-..."

Chat Completions

from openai import OpenAI

client = OpenAI()  # tự đọc OPENAI_API_KEY từ env

# Basic call
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": "Bạn là assistant lập trình Python."},
        {"role": "user", "content": "Giải thích list comprehension với ví dụ"}
    ],
    max_tokens=500,
    temperature=0.3
)
print(response.choices[0].message.content)
print(f"Tokens dùng: {response.usage.total_tokens}")

Streaming

# Streaming — in ra từng token ngay khi nhận được
with client.chat.completions.stream(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Viết bài thơ về Hà Nội"}]
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)
print()  # Newline cuối

Structured Output với Pydantic

from pydantic import BaseModel
from typing import List

class CodeReview(BaseModel):
    bugs: List[str]
    improvements: List[str]
    security_issues: List[str]
    overall_score: int  # 1-10

response = client.beta.chat.completions.parse(
    model="gpt-4o",
    messages=[
        {"role": "user", "content": """Review code này:
```python
import os
password = os.environ.get('PASSWORD', 'admin123')
query = f"SELECT * FROM users WHERE id = {user_id}"
```"""}
    ],
    response_format=CodeReview
)
review = response.choices[0].message.parsed
print(f"Bugs: {review.bugs}")
print(f"Security: {review.security_issues}")
print(f"Score: {review.overall_score}/10")

Vision (Image Input)

import base64
from pathlib import Path

def analyze_image(image_path: str, question: str) -> str:
    image_data = base64.b64encode(Path(image_path).read_bytes()).decode()
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{
            "role": "user",
            "content": [
                {"type": "text", "text": question},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/jpeg;base64,{image_data}",
                        "detail": "high"  # "low" tốn ít tokens hơn
                    }
                }
            ]
        }]
    )
    return response.choices[0].message.content

# Hoặc dùng URL trực tiếp
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Mô tả ảnh này"},
            {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}}
        ]
    }]
)

---

3. Anthropic Claude API

Setup

pip install anthropic
export ANTHROPIC_API_KEY="sk-ant-..."

Messages API

import anthropic

client = anthropic.Anthropic()

# Basic
message = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    system="Bạn là expert về distributed systems.",
    messages=[
        {"role": "user", "content": "Giải thích CAP theorem"}
    ]
)
print(message.content[0].text)
print(f"Input tokens: {message.usage.input_tokens}")
print(f"Output tokens: {message.usage.output_tokens}")

Streaming

with client.messages.stream(
    model="claude-opus-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Viết class Python cho binary search tree"}]
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Extended Thinking (Claude 3.7+)

# Claude có thể "suy nghĩ" trước khi trả lời (như chain-of-thought nội bộ)
response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000  # Tokens dành cho thinking
    },
    messages=[{
        "role": "user",
        "content": "Giải bài toán: 17 × 19 - sqrt(144) + 2^8"
    }]
)

for block in response.content:
    if block.type == "thinking":
        print(f"[Thinking]: {block.thinking[:200]}...")
    elif block.type == "text":
        print(f"[Answer]: {block.text}")

Long Document Analysis

# Claude hỗ trợ 200K context — phân tích tài liệu dài
with open("contract.txt") as f:
    document = f.read()

response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=2048,
    messages=[{
        "role": "user",
        "content": f"""Phân tích hợp đồng sau và trả lời:
1. Các điều khoản quan trọng nhất
2. Rủi ro tiềm ẩn
3. Ngày hết hạn (nếu có)

<document>
{document}
</document>"""
    }]
)

---

4. Google Gemini API

Setup

pip install google-generativeai
export GOOGLE_API_KEY="AIza..."

Basic Usage

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GOOGLE_API_KEY"])

model = genai.GenerativeModel(
    model_name="gemini-2.0-flash",
    system_instruction="Bạn là tutor Python cho người mới học."
)

response = model.generate_content("Decorators trong Python là gì?")
print(response.text)

Multi-turn Chat

chat = model.start_chat(history=[])

while True:
    user_input = input("Bạn: ")
    if user_input.lower() == "quit":
        break

    response = chat.send_message(user_input)
    print(f"Gemini: {response.text}")

Multimodal (Image + Video + Audio)

import PIL.Image

# Image analysis
image = PIL.Image.open("chart.png")
response = model.generate_content([
    "Phân tích biểu đồ này và đưa ra insights",
    image
])
print(response.text)

# Gemini 1.5 Pro — 1M context với video
video_file = genai.upload_file("presentation.mp4")
response = model.generate_content([
    "Tóm tắt nội dung video này",
    video_file
])

---

5. Unified Client Pattern

from abc import ABC, abstractmethod
from dataclasses import dataclass
from typing import Iterator

@dataclass
class LLMResponse:
    content: str
    input_tokens: int
    output_tokens: int

class LLMClient(ABC):
    @abstractmethod
    def complete(self, prompt: str, system: str = "") -> LLMResponse:
        pass

    @abstractmethod
    def stream(self, prompt: str, system: str = "") -> Iterator[str]:
        pass

class OpenAIClient(LLMClient):
    def __init__(self, model="gpt-4o"):
        from openai import OpenAI
        self.client = OpenAI()
        self.model = model

    def complete(self, prompt: str, system: str = "") -> LLMResponse:
        messages = []
        if system:
            messages.append({"role": "system", "content": system})
        messages.append({"role": "user", "content": prompt})

        r = self.client.chat.completions.create(model=self.model, messages=messages)
        return LLMResponse(
            content=r.choices[0].message.content,
            input_tokens=r.usage.prompt_tokens,
            output_tokens=r.usage.completion_tokens
        )

    def stream(self, prompt: str, system: str = "") -> Iterator[str]:
        messages = [{"role": "user", "content": prompt}]
        with self.client.chat.completions.stream(
            model=self.model, messages=messages
        ) as s:
            yield from s.text_stream

class AnthropicClient(LLMClient):
    def __init__(self, model="claude-opus-4-5"):
        import anthropic
        self.client = anthropic.Anthropic()
        self.model = model

    def complete(self, prompt: str, system: str = "") -> LLMResponse:
        r = self.client.messages.create(
            model=self.model, max_tokens=2048,
            system=system or "You are a helpful assistant.",
            messages=[{"role": "user", "content": prompt}]
        )
        return LLMResponse(
            content=r.content[0].text,
            input_tokens=r.usage.input_tokens,
            output_tokens=r.usage.output_tokens
        )

    def stream(self, prompt: str, system: str = "") -> Iterator[str]:
        with self.client.messages.stream(
            model=self.model, max_tokens=2048,
            messages=[{"role": "user", "content": prompt}]
        ) as s:
            yield from s.text_stream

# Dùng
def get_client(provider: str = "openai") -> LLMClient:
    if provider == "openai":
        return OpenAIClient()
    elif provider == "anthropic":
        return AnthropicClient()
    raise ValueError(f"Unknown provider: {provider}")

llm = get_client("openai")
response = llm.complete("Giải thích REST API cho người mới")
print(response.content)

---

6. Cost Optimization

# Pricing tham khảo (USD per 1M tokens, 2025)
PRICING = {
    "gpt-4o":           {"input": 2.50,  "output": 10.00},
    "gpt-4o-mini":      {"input": 0.15,  "output": 0.60},
    "claude-opus-4-5":  {"input": 3.00,  "output": 15.00},
    "claude-haiku-4-5": {"input": 0.80,  "output": 4.00},
    "gemini-2.0-flash": {"input": 0.10,  "output": 0.40},
}

def estimate_cost(model: str, input_tokens: int, output_tokens: int) -> float:
    p = PRICING.get(model, {"input": 1.0, "output": 1.0})
    return (input_tokens * p["input"] + output_tokens * p["output"]) / 1_000_000

# Chiến lược tiết kiệm
tips = """
1. Model routing: dùng cheap model (gpt-4o-mini) cho tasks đơn giản,
   expensive model chỉ khi cần

2. Prompt caching: Anthropic cache system prompts dài
   → tiết kiệm 90% input cost cho repeated calls

3. Batch API: OpenAI Batch API 50% cheaper (async, 24h delay)

4. Context window management: trim conversation history
   → tránh truyền lại toàn bộ history mỗi call

5. Response caching: cache kết quả cho cùng input
   (đặc biệt hữu ích cho RAG retrieval)
"""

---

Tổng kết

Task	Gợi ý
Coding, agents	GPT-4o hoặc Claude Opus
Cheap, fast	GPT-4o-mini, Claude Haiku, Gemini Flash
Long document (100K+)	Claude Opus (200K), Gemini Pro (1M)
Multimodal	GPT-4o, Gemini, Claude (images)
Structured output	OpenAI JSON mode + Pydantic
Streaming UI	Tất cả đều hỗ trợ

Bài tiếp theo (cuối series): Deploying LLMs tự host với Ollama và vLLM, evaluation và production checklist.