प्रति चरण टोकन बजट

response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=messages,
    max_tokens=1024,   # cap per response
)

STEP_BUDGETS = {
    'planner': 512,
    'classifier': 64,
    'final_synthesis': 2048,
    'tool_call': 256
}

response = client.chat.completions.create(
    model=model,
    messages=messages,
    max_tokens=STEP_BUDGETS[step_name]
)

import tiktoken
enc = tiktoken.encoding_for_model('gpt-4o-mini')

def trim_to_budget(text, max_tokens=4000):
    toks = enc.encode(text)
    if len(toks) > max_tokens:
        return enc.decode(toks[:max_tokens])
    return text

MAX_TOKENS_PER_RUN = 50_000
total = 0
for step in agent_steps:
    r = call_step(step)
    total += r.usage.total_tokens
    if total > MAX_TOKENS_PER_RUN:
        return 'Budget exhausted; partial result'

PRICES = {'gpt-4o-mini': {'in': 0.15e-6, 'out': 0.60e-6}}

for step in agent_steps:
    r = call_step(step)
    cost = r.usage.prompt_tokens * PRICES[model]['in'] + r.usage.completion_tokens * PRICES[model]['out']
    total_cost += cost
    if total_cost > MAX_COST_PER_RUN:
        break

if step_difficulty == 'easy':
    max_tokens = 256
elif step_difficulty == 'hard':
    max_tokens = 2048

preview = llm.invoke(f'How many tokens (approx) does {question} need to answer?').content
budget = parse_int(preview, default=512)
final = llm.invoke(question, max_tokens=budget)

tool_result = run_tool(...)
text = json.dumps(tool_result)
if len(text) > 4000:
    text = text[:4000] + '\n...[truncated]'

if total_message_tokens(messages) > 8000:
    messages = compact(messages)

metrics.observe('step.planner.tokens', r.usage.total_tokens, tags={'agent': 'qa'})