This model's maximum context length is exceeded

1
2
3
4
5
6

# NG: 長すぎる会話履歴
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": very_long_text},  # 数千トークン
    # ... 多数のメッセージ
]

1
2
3
4
5
6

# NG: 入力 + max_tokens がコンテキスト長を超える
response = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=messages,  # 3500トークン
    max_tokens=1000     # 合計4500 > 4096
)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

import tiktoken

def count_tokens(messages, model="gpt-4"):
    encoding = tiktoken.encoding_for_model(model)
    num_tokens = 0
    for message in messages:
        num_tokens += 4  # メッセージのオーバーヘッド
        for key, value in message.items():
            num_tokens += len(encoding.encode(value))
    num_tokens += 2  # 返答のプライミング
    return num_tokens

# 使用例
token_count = count_tokens(messages)
print(f"Token count: {token_count}")

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

def trim_messages(messages, max_tokens=3000, model="gpt-4"):
    encoding = tiktoken.encoding_for_model(model)

    # システムメッセージは保持
    system_message = messages[0] if messages[0]["role"] == "system" else None
    other_messages = messages[1:] if system_message else messages

    # 新しいメッセージから優先的に保持
    trimmed = []
    total_tokens = 0

    for msg in reversed(other_messages):
        msg_tokens = len(encoding.encode(msg["content"]))
        if total_tokens + msg_tokens <= max_tokens:
            trimmed.insert(0, msg)
            total_tokens += msg_tokens
        else:
            break

    if system_message:
        trimmed.insert(0, system_message)

    return trimmed

1
2
3
4
5
6

# gpt-4 (8K) → gpt-4-turbo (128K)
response = client.chat.completions.create(
    model="gpt-4-turbo",  # 128Kコンテキスト
    messages=messages,
    max_tokens=4096
)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

def summarize_if_needed(text, max_length=2000):
    if len(text) <= max_length:
        return text

    # 長いテキストは要約
    response = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[
            {"role": "system", "content": "Summarize the following text concisely."},
            {"role": "user", "content": text[:4000]}  # 最初の部分だけ
        ],
        max_tokens=500
    )
    return response.choices[0].message.content

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17

import tiktoken

def safe_completion(messages, model="gpt-4", desired_max_tokens=1000):
    encoding = tiktoken.encoding_for_model(model)
    input_tokens = count_tokens(messages, model)

    # モデルの最大コンテキスト長
    max_context = {"gpt-4": 8192, "gpt-4-turbo": 128000, "gpt-3.5-turbo": 16385}

    available_tokens = max_context.get(model, 4096) - input_tokens
    safe_max_tokens = min(desired_max_tokens, available_tokens - 100)

    return client.chat.completions.create(
        model=model,
        messages=messages,
        max_tokens=max(safe_max_tokens, 100)
    )

1
2
3
4
5
6
7

from langchain.memory import ConversationSummaryBufferMemory

# トークン数に基づいて自動的に要約
memory = ConversationSummaryBufferMemory(
    llm=llm,
    max_token_limit=2000
)