-
Checked other resources
Commit to Help
Example Codefrom enum import Enum
import ast
import json
from langchain_core.output_parsers import JsonOutputParser
from langchain_core.prompts import PromptTemplate
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain_community.llms import Ollama
model = Ollama(base_url='http://155.198.89.235:11434',model="llama3")
class Prediction(Enum):
increase = "Increase"
decrease = "Decrease"
nochange = "No Change"
class Article(BaseModel):
prediction: Prediction = Field(description = "What can be inferred about natural gas prices from the article? increase, decrease or no change? Do not add square brackets.")
year: str = Field(description = "An estimate of when these changes will happen.")
year = NGID['_id'][:4]
query = 'The article headline is "' + title + '". The article subtitle is "' + subtitle + '". The body of the article is as follows: "' + body + '". The article was published on ' + date + '.'
parser = JsonOutputParser(pydantic_object=Article)
prompt = PromptTemplate(
template = "Answer the user query in python dictionary format only.\n{format_instructions}\n{query}\n",
input_variables = ["query"],
partial_variables = {"format_instructions": parser.get_format_instructions()},
)
chain = prompt | model | parser
dic = chain.invoke({"query": query})DescriptionI'm trying to use the Json output parser to give me a prediction and year, but the json output parser sometimes returns outputs that give extra information or its just not in dictionary format, e.g. the error below: langchain_core.exceptions.OutputParserException: Invalid json output: Here is the answer in Python dictionary format: Is there a way to use the output-fixing parser or the retry parser with the json parser? System InfoSystem Information
Package Information
Packages not installed (Not Necessarily a Problem)The following packages were not found:
|
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 19 replies
-
|
To handle issues with extra information or incorrect dictionary formats in the output using the
Here is an example code snippet demonstrating this: from enum import Enum
from langchain_core.output_parsers import JsonOutputParser
from langchain.output_parsers.fix import OutputFixingParser
from langchain.output_parsers.retry import RetryOutputParser
from langchain_core.prompts import PromptTemplate
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain_community.llms import Ollama
model = Ollama(base_url='http://155.198.89.235:11434', model="llama3")
class Prediction(Enum):
increase = "Increase"
decrease = "Decrease"
nochange = "No Change"
class Article(BaseModel):
prediction: Prediction = Field(description="What can be inferred about natural gas prices from the article? increase, decrease or no change? Do not add square brackets.")
year: str = Field(description="An estimate of when these changes will happen.")
year = NGID['_id'][:4]
query = 'The article headline is "' + title + '". The article subtitle is "' + subtitle + '". The body of the article is as follows: "' + body + '". The article was published on ' + date + '.'
# Step 1: Create an instance of JsonOutputParser
json_parser = JsonOutputParser(pydantic_object=Article)
# Step 2: Create an instance of OutputFixingParser using the from_llm method
output_fixing_parser = OutputFixingParser.from_llm(
llm=model,
parser=json_parser,
max_retries=3 # You can set the maximum number of retries
)
# Alternatively, create an instance of RetryOutputParser using the from_llm method
retry_parser = RetryOutputParser.from_llm(
llm=model,
parser=json_parser,
max_retries=3 # You can set the maximum number of retries
)
prompt = PromptTemplate(
template="Answer the user query in python dictionary format only.\n{format_instructions}\n{query}\n",
input_variables=["query"],
partial_variables={"format_instructions": json_parser.get_format_instructions()},
)
chain = prompt | model | output_fixing_parser # or use retry_parser
dic = chain.invoke({"query": query})
print(dic)In this example:
This setup will help handle issues with extra information or incorrect dictionary formats in the output by retrying the parsing process using the language model [1][2][3][4].
|
Beta Was this translation helpful? Give feedback.
-
|
@dosu it still sometimes generates outputs not in the desired format. Is there a way I can check if it only contains the keys 'pricechange' and 'year', and if it doesnt it will try to parse again? |
Beta Was this translation helpful? Give feedback.
-
|
To ensure the output only contains the keys
Here is the updated code: from enum import Enum
from langchain_core.output_parsers import JsonOutputParser
from langchain.output_parsers.retry import RetryOutputParser, OutputParserException
from langchain_core.prompts import PromptTemplate
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain_community.llms import Ollama
from langchain_core.prompt_values import StringPromptValue
model = Ollama(base_url='http://155.198.89.235:11434', model="llama3")
class PriceChange(Enum):
increase = "Increase"
decrease = "Decrease"
nochange = "No Change"
class Article(BaseModel):
pricechange: PriceChange = Field(description="What can be inferred about natural gas prices from the article? increase, decrease or no change? Do not add square brackets and do not give numbers.")
year: str = Field(description="An estimate of the year when these changes will happen.")
# Create a JsonOutputParser instance
json_parser = JsonOutputParser(pydantic_object=Article)
# Create a RetryOutputParser instance using the from_llm class method
retry_parser = RetryOutputParser.from_llm(
llm=model,
parser=json_parser,
prompt=PromptTemplate.from_template("There should only be 2 fields in the json, pricechange and year."),
max_retries=3
)
# Modify the parse_with_prompt method to check for required keys
class CustomRetryOutputParser(RetryOutputParser):
def parse_with_prompt(self, completion: str, prompt_value: StringPromptValue) -> dict:
retries = 0
while retries <= self.max_retries:
try:
result = self.parser.parse(completion)
if 'pricechange' in result and 'year' in result:
return result
else:
raise OutputParserException("Missing required keys")
except OutputParserException as e:
if retries == self.max_retries:
raise e
else:
retries += 1
completion = self.retry_chain.run(
instructions=self.parser.get_format_instructions(),
completion=completion,
error=repr(e),
)
raise OutputParserException("Failed to parse")
# Use the custom parser
custom_retry_parser = CustomRetryOutputParser.from_llm(
llm=model,
parser=json_parser,
prompt=PromptTemplate.from_template("There should only be 2 fields in the json, pricechange and year."),
max_retries=3
)
article = 'The article headline is "' + title + '". The article subtitle is "' + subtitle + '". The body of the article is as follows: "' + body + '". The article was published on ' + date + '.'
prompt = PromptTemplate(
template="""What can be inferred about natural gas prices from the article and what year might the change happen?
Respond with a JSON object containing only the fields 'pricechange' and 'year'.
{format_instructions}
Article: {article}""",
input_variables=["article"],
partial_variables={"format_instructions": json_parser.get_format_instructions()},
)
# Convert the prompt to a PromptValue
prompt_value = prompt.format_prompt(article=article)
# Ensure the prompt_value is passed as a list of strings
llm_result = model.generate([prompt_value.to_string()]) # Assuming model.generate() returns an LLMResult
# Extract the text from the LLMResult object
completion = llm_result.generations[0][0].text
# Use the parse_with_prompt method
parsed_output = custom_retry_parser.parse_with_prompt(completion, prompt_value)
print(parsed_output)Key Changes:
This modification ensures that the parser checks for the presence of the required keys and retries the parsing process if they are missing [1][2][3].
|
Beta Was this translation helpful? Give feedback.
-
|
@dosu is there a way to make the check for the keys not case sensitive? |
Beta Was this translation helpful? Give feedback.
-
|
To make the check for the keys from enum import Enum
from langchain_core.output_parsers import JsonOutputParser
from langchain.output_parsers.retry import RetryOutputParser, OutputParserException
from langchain_core.prompts import PromptTemplate
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain_community.llms import Ollama
from langchain_core.prompt_values import StringPromptValue
model = Ollama(base_url='http://155.198.89.235:11434', model="llama3")
class PriceChange(Enum):
increase = "Increase"
decrease = "Decrease"
nochange = "No Change"
class Article(BaseModel):
pricechange: PriceChange = Field(description="What can be inferred about natural gas prices from the article? increase, decrease or no change? Do not add square brackets and do not give numbers.")
year: str = Field(description="An estimate of the year when these changes will happen.")
# Create a JsonOutputParser instance
json_parser = JsonOutputParser(pydantic_object=Article)
# Modify the parse_with_prompt method to check for required keys case-insensitively
class CustomRetryOutputParser(RetryOutputParser):
def parse_with_prompt(self, completion: str, prompt_value: StringPromptValue) -> dict:
retries = 0
while retries <= self.max_retries:
try:
result = self.parser.parse(completion)
# Convert keys to lowercase for case-insensitive check
result_lower = {k.lower(): v for k, v in result.items()}
if 'pricechange' in result_lower and 'year' in result_lower:
return result
else:
raise OutputParserException("Missing required keys")
except OutputParserException as e:
if retries == self.max_retries:
raise e
else:
retries += 1
completion = self.retry_chain.run(
instructions=self.parser.get_format_instructions(),
completion=completion,
error=repr(e),
)
raise OutputParserException("Failed to parse")
# Use the custom parser
custom_retry_parser = CustomRetryOutputParser.from_llm(
llm=model,
parser=json_parser,
prompt=PromptTemplate.from_template("There should only be 2 fields in the json, pricechange and year."),
max_retries=3
)
article = 'The article headline is "' + title + '". The article subtitle is "' + subtitle + '". The body of the article is as follows: "' + body + '". The article was published on ' + date + '.'
prompt = PromptTemplate(
template="""What can be inferred about natural gas prices from the article and what year might the change happen?
Respond with a JSON object containing only the fields 'pricechange' and 'year'.
{format_instructions}
Article: {article}""",
input_variables=["article"],
partial_variables={"format_instructions": json_parser.get_format_instructions()},
)
# Convert the prompt to a PromptValue
prompt_value = prompt.format_prompt(article=article)
# Ensure the prompt_value is passed as a list of strings
llm_result = model.generate([prompt_value.to_string()]) # Assuming model.generate() returns an LLMResult
# Extract the text from the LLMResult object
completion = llm_result.generations[0][0].text
# Use the parse_with_prompt method
parsed_output = custom_retry_parser.parse_with_prompt(completion, prompt_value)
print(parsed_output)In this modified
|
Beta Was this translation helpful? Give feedback.
-
|
This is not an issue with the JSONParser, but with the LLM itself. You can filter out the output. Also you may get better performance by forcing the model to use JSON mode. |
Beta Was this translation helpful? Give feedback.
-
|
How to force llm to use JSON mode |
Beta Was this translation helpful? Give feedback.
To make the check for the keys
pricechangeandyearcase-insensitive in theCustomRetryOutputParser, you can modify theparse_with_promptmethod to convert the keys to lowercase before performing the check. Here is an example of how you can achieve this: