Usar ferramentas externas com generate_text

Ver em ai.google.dev Testar um bloco do Colab Veja o notebook no GitHub

Em alguns casos de uso, é possível interromper a geração de um modelo para inserir resultados específicos. Por exemplo, modelos de linguagem podem ter problemas com problemas aritméticos complicados, como problemas com palavras. Neste tutorial, mostramos um exemplo de uso de uma ferramenta externa com o método genai.generate_text para gerar a resposta correta para um problema matemático.

Este exemplo específico usa a ferramenta numexpr para realizar a aritmética, mas é possível usar o mesmo procedimento para integrar outras ferramentas específicas ao seu caso de uso. Confira a seguir um resumo das etapas:

  1. Determine uma tag start e end para demarcar o texto a ser enviado à ferramenta.
  2. Crie um comando instruindo o modelo a usar as tags no resultado.
  3. Inclua a tag end na das stop_sequences transmitidas para generate_text.
  4. No resultado do modelo, use o texto entre as tags start e end como entrada para a ferramenta.
  5. Execute a ferramenta e adicione a saída ao comando.
  6. Chame generate_text novamente para que o modelo continue com a saída da ferramenta.

Configuração

pip install -q google.generativeai
import google.generativeai as genai
genai.configure(api_key='YOUR API KEY')

from google.api_core import retry

@retry.Retry()
def generate_text(*args, **kwargs):
  return genai.generate_text(*args, **kwargs)
models = [m for m in genai.list_models() if 'generateText' in m.supported_generation_methods]
model = models[0].name
print(model)
models/text-bison-001

Tentar resolver o problema diretamente

Aqui está o problema de palavras que você vai resolver:

question = """
I have 77 houses, each with 31 cats.
Each cat owns 14 mittens, and 6 hats.
Each mitten was knit from 141m of yarn, each hat from 55m.
How much yarn was needed to make all the items?
"""
prompt_template = """
You are an expert at solving word problems. Here's one:

{question}

Work through it step by step, and show your work.
One step per line.

Your solution:
"""

Tente como está:

completion = generate_text(
    model=model,
    prompt=prompt_template.format(question=question),
    # The maximum length of the response
    max_output_tokens=800,
)

print(completion.result)
In the houses there are 77 * 31 = 2387 cats.
So they need 2387 * 14 = 33418 mittens.
And they need 2387 * 6 = 14322 hats.
In total they need 33418 * 141 + 14322 * 55 = 5554525m of yarn.
The answer: 5554525.

Normalmente, o comando como está retorna um resultado incorreto. Em geral, ele acerta as etapas, mas a aritmética está errada.

A resposta deve ser:

answer = 77*31*14*141 + 77*31*6*55
answer
5499648

Dizer ao modelo para usar uma calculadora

Na próxima tentativa, forneça ao modelo instruções sobre como acessar a calculadora. Para fazer isso, especifique as tags start e end que o modelo pode usar para indicar onde um cálculo é necessário. Adicione algo como o seguinte ao comando:

calc_prompt_template = """
You are an expert at solving word problems. Here's a question:

{question}

-------------------

When solving this problem, use the calculator for any arithmetic.

To use the calculator, put an expression between <calc></calc> tags.
The answer will be printed after the </calc> tag.

For example: 2 houses  * 8 cats/house = <calc>2 * 8</calc> = 16 cats

-------------------

Work through it step by step, and show your work.
One step per line.

Your solution:
"""

calc_prompt = calc_prompt_template.format(question=question)

Para dar ao modelo acesso à saída dessa "calculadora", você vai precisar pausar a geração e inserir o resultado. Use o argumento stop_sequences para parar na tag </calc>:

completion = generate_text(
    model=model,
    prompt=calc_prompt,
    stop_sequences=["</calc>"],
    # The maximum length of the response
    max_output_tokens=800,
    candidate_count=1,
)

result = completion.result
print(result)
In each house, there are <calc>31 * 14

O stop_sequence não está incluído no resultado. Divida a expressão, execute-a pela calculadora e adicione de volta ao resultado:

# Use re to clear units from the calculator expressions
import re
# Use numexpr since `eval` is unsafe.
import numexpr


def calculator(result):
  result, expression = result.rsplit('<calc>', 1)

  # Strip any units like "cats / house"
  clean_expression = re.sub("[a-zA-Z]([ /a-zA-Z]*[a-zA-Z])?",'', expression)

  # `eval` is unsafe use numexpr
  result = f"{result}<calc>{expression}</calc> = {str(numexpr.evaluate(clean_expression))}"
  return result
print(calculator(result))
In each house, there are <calc>31 * 14</calc> = 434

Agora, anexe isso ao prompt e execute o modelo novamente para que ele possa continuar de onde parou:

continue_prompt=calc_prompt +"\n"+ "-"*80 + "\n" + calculator(result)

completion = generate_text(
    model=model,
    prompt=continue_prompt,
    stop_sequences=["</calc>"],
    # The maximum length of the response
    max_output_tokens=800,
    candidate_count=1,
)

print(completion.result)
mittens.
In each house, there are <calc>31 * 6

Desta vez, o modelo continuou o texto do último cálculo e passou para o próximo. Agora, execute-o em loop para resolver totalmente o problema da palavra:

def solve(question=question):
  results = []

  for n in range(10):
    prompt = calc_prompt_template.format(question=question)

    prompt += " ".join(results)

    completion = generate_text(
        model=model,
        prompt=prompt,
        stop_sequences=["</calc>"],
        # The maximum length of the response
        max_output_tokens=800,
    )

    result = completion.result
    if '<calc>' in result:
      result = calculator(result)

    results.append(result)
    print('-'*40)
    print(result)
    if str(answer) in result:
      break
    if "<calc>" not in  result:
      break

  is_good = any(str(answer) in r for r in results)

  print("*"*100)
  if is_good:
    print("Success!")
  else:
    print("Failure!")
  print("*"*100)

  return is_good
solve(question);
----------------------------------------
The total number of cats is <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
The total number of mittens is <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens.
The total amount of yarn needed for the mittens is <calc>33418 * 141</calc> = 4711938
----------------------------------------
m.
The total number of hats is <calc>2387 * 6</calc> = 14322
----------------------------------------
hats.
 The total amount of yarn needed for the hats is <calc>14322 * 55</calc> = 787710
----------------------------------------
m.
In total, <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!

É possível fazer isso algumas vezes para estimar a taxa de resolução:

good = []

for n in range(10):
  good.append(solve(question))
----------------------------------------
There are <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
They need <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens.
The mittens need <calc>33418 * 141</calc> = 4711938
----------------------------------------
m of yarn.
They need <calc>2387 * 6</calc> = 14322
----------------------------------------
hats.
The hats need <calc>14322 * 55</calc> = 787710
----------------------------------------
m of yarn.
 They need a total of <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
There are <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
So for the mittens, we need <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens.
That means we need <calc>33418 * 141</calc> = 4711938
----------------------------------------
m of yarn for mittens.
For the hats, we need <calc>2387 * 6</calc> = 14322
----------------------------------------
hats.
That means we need <calc>14322 * 55</calc> = 787710
----------------------------------------
m of yarn for hats.
 In total we need <calc>787710 + 4711938</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
In the 77 houses I have <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
They need <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens.
The mittens need <calc>33418 * 141</calc> = 4711938
----------------------------------------
m of yarn.
They need <calc>2387 * 6</calc> = 14322
----------------------------------------
hats.
The hats need <calc>14322 * 55</calc> = 787710
----------------------------------------
m of yarn.
 So, in total I need <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
The number of cats is <calc>77 * 31</calc> = 2387
----------------------------------------
. Each cat needs <calc>14 * 141</calc> = 1974
----------------------------------------
m of yarn for mittens. So we need <calc>1974 * 2387</calc> = 4711938
----------------------------------------
m of yarn for mittens. Each cat needs <calc>6 * 55</calc> = 330
----------------------------------------
m of yarn for hats. So we need <calc>330 * 2387</calc> = 787710
----------------------------------------
m of yarn for hats. So in total we need <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
There are <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
Each cat needs <calc>14 * 141</calc> = 1974
----------------------------------------
yarn for mittens.
All cats need <calc>2387 * 1974</calc> = 4711938
----------------------------------------
yarn for mittens.
Each cat needs <calc>6 * 55</calc> = 330
----------------------------------------
yarn for hats.
All cats need <calc>2387 * 330</calc> = 787710
----------------------------------------
yarn for hats.
 All in all, you need <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
There are <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
Each cat needs <calc>14 + 6</calc> = 20
----------------------------------------
items.
So we need <calc>20 * 2387</calc> = 47740
----------------------------------------
items in total.
Each mitten needs <calc>141</calc> = 141
----------------------------------------
m of yarn.
So all the mittens need <calc>141 * 47740</calc> = 6731340
----------------------------------------
m of yarn.
 Each hat needs <calc>55</calc> = 55
----------------------------------------
m of yarn.
So all the hats need <calc>55 * 47740</calc> = 2625700
----------------------------------------
m of yarn.
 In total, we need <calc>6731340 + 2625700</calc> = 9357040
----------------------------------------
m of yarn. There are <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
Each cat needs <calc>14 + 6</calc> = 20
********************************************************************************
Failure!
----------------------------------------
There are <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
There are <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens.
There are <calc>2387 * 6</calc> = 14322
----------------------------------------
hats.
There was <calc>141 * 33418</calc> = 4711938
----------------------------------------
m of yarn needed for mittens.
There was <calc>55 * 14322</calc> = 787710
----------------------------------------
m of yarn needed for hats.
 So there was <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
There are <calc>77 * 31</calc> = 2387
----------------------------------------
cats in total. 
They need <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens. 
That's <calc>33418 * 141</calc> = 4711938
----------------------------------------
meters of yarn for mittens. 
They need <calc>2387 * 6</calc> = 14322
----------------------------------------
hats. 
That's <calc>14322 * 55</calc> = 787710
----------------------------------------
meters of yarn for hats. 
So, they need <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
There are 77 houses * 31 cats / house = <calc>77 * 31</calc> = 2387
----------------------------------------
cats.
So we need <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens.
So we need <calc>33418 * 141</calc> = 4711938
----------------------------------------
m of yarn for mittens.
So we need <calc>2387 * 6</calc> = 14322
----------------------------------------
hats.
 So we need <calc>14322 * 55</calc> = 787710
----------------------------------------
m of yarn for hats.
In total, we need <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
----------------------------------------
In total there are 77 houses * 31 cats / house = <calc>77 * 31</calc> = 2387
----------------------------------------
cats. In total 2387 cats * 14 mittens / cat = <calc>2387 * 14</calc> = 33418
----------------------------------------
mittens. In total 33418 mittens * 141m / mitten = <calc>33418 * 141</calc> = 4711938
----------------------------------------
m of yarn for mittens. In total 2387 cats * 6 hats / cat = <calc>2387 * 6</calc> = 14322
----------------------------------------
hats. In total 14322 hats * 55m / hat = <calc>14322 * 55</calc> = 787710
----------------------------------------
m of yarn for hats. In total we need 4711938 m of yarn for mittens + 787710 m of yarn for hats = <calc>4711938 + 787710</calc> = 5499648
********************************************************************************
Success!
import numpy as np
np.mean(good)
0.9