Best Practices

# BEFORE (~85 tokens):
"Hi there! I would really appreciate your help with
something. Could you please take a look at the text
below and provide me with a brief summary of the main
points? I'd like the summary to be concise and capture
the key ideas. Thank you so much in advance!"

# AFTER (~10 tokens):
"Summarize the key points of this text:"

# Same quality output, ~88% fewer input tokens

# VERBOSE system prompt (~120 tokens):
"You are a professional software developer who specializes
in writing Python code. When the user asks you to write code,
always include type hints, docstrings, and follow PEP 8
coding standards. Handle edge cases and include error handling.
Format all code using markdown code blocks with the Python
language identifier."

# OPTIMIZED system prompt (~35 tokens):
"Expert Python dev. Always: type hints, docstrings, PEP 8,
error handling, edge cases. Use ```python code blocks."

import tiktoken

def trim_messages(messages, max_tokens=4000, model="gpt-4o"):
    """Keep the most recent messages within the token limit."""
    enc = tiktoken.encoding_for_model(model)
    total = 0
    trimmed = []

    # Always keep system message
    system_msg = messages[0] if messages[0]["role"] == "system" else None
    if system_msg:
        total += len(enc.encode(system_msg["content"])) + 4
        trimmed.append(system_msg)

    # Add messages from newest to oldest
    user_msgs = messages[1:] if system_msg else messages
    for msg in reversed(user_msgs):
        msg_tokens = len(enc.encode(msg["content"])) + 4
        if total + msg_tokens > max_tokens:
            break
        total += msg_tokens
        trimmed.insert(1 if system_msg else 0, msg)

    return trimmed

# JSON format (~45 tokens):
{"users": [{"name": "Alice", "role": "admin"}, {"name": "Bob", "role": "user"}]}

# CSV format (~15 tokens):
name,role
Alice,admin
Bob,user

# Plain text (~12 tokens):
Alice (admin), Bob (user)

import logging
from datetime import datetime

logger = logging.getLogger("token_usage")

def log_usage(response, model, endpoint):
    usage = response.usage
    cost = calculate_cost(model, usage.prompt_tokens, usage.completion_tokens)

    logger.info(
        "API Call | model=%s | input=%d | output=%d | "
        "total=%d | cost=$%.4f | endpoint=%s | time=%s",
        model,
        usage.prompt_tokens,
        usage.completion_tokens,
        usage.total_tokens,
        cost,
        endpoint,
        datetime.now().isoformat(),
    )

    # Alert on unusually high token counts
    if usage.total_tokens > 10000:
        logger.warning(
            "High token usage detected: %d tokens ($%.4f)",
            usage.total_tokens,
            cost,
        )