Output Parsers

from langchain_core.output_parsers import StrOutputParser, JsonOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, Field
from typing import List, Literal

llm = ChatOpenAI(model='gpt-4o-mini', temperature=0)

# 1) StrOutputParser — just the text content
chain = ChatPromptTemplate.from_template('Translate to French: {text}') | llm | StrOutputParser()
chain.invoke({'text': 'hello'})    # 'bonjour'
# Without a parser you'd get an AIMessage object — annoying for pipelines.

# 2) Pydantic — structured, validated output
class Recipe(BaseModel):
    title: str        = Field(description='dish name')
    servings: int     = Field(description='number of servings', ge=1)
    ingredients: List[str]
    steps: List[str]
    cuisine: Literal['italian', 'french', 'thai', 'mexican', 'other']

parser = PydanticOutputParser(pydantic_object=Recipe)

prompt = ChatPromptTemplate.from_messages([
    ('system', 'Return a recipe as JSON.\n{format_instructions}'),
    ('user', '{request}'),
]).partial(format_instructions=parser.get_format_instructions())

chain = prompt | llm | parser
recipe = chain.invoke({'request': 'A quick pasta dinner'})
print(type(recipe))                  # <class 'Recipe'>
print(recipe.title, recipe.servings) # validated, typed

# 3) JsonOutputParser — just a dict, no schema enforcement
parser = JsonOutputParser()
chain  = prompt | llm | parser
data   = chain.invoke({'request': '…'})  # dict

# JsonOutputParser with Pydantic for schema docs
parser = JsonOutputParser(pydantic_object=Recipe)

# 4) Modern: with_structured_output (recommended for new code)
structured_llm = llm.with_structured_output(Recipe)
recipe = structured_llm.invoke('A quick pasta dinner')
# No prompt instructions or parser needed — uses tool calling under the hood.

# 5) Streaming JSON — JsonOutputParser streams partial dicts
chain = prompt | llm | JsonOutputParser()
for partial in chain.stream({'request': '…'}):
    print(partial)  # {} → {'title': 'Pasta'} → {'title': 'Pasta', 'servings': 4} …

# 6) OutputFixingParser — wrap a parser so failures retry through the LLM
from langchain.output_parsers import OutputFixingParser

fixing = OutputFixingParser.from_llm(parser=parser, llm=llm)
try:
    recipe = fixing.parse(messy_text)
except Exception as e:
    log.error('still failed after retry', exc_info=e)

# 7) RetryOutputParser — re-runs original prompt with the bad output as context
from langchain.output_parsers import RetryOutputParser
retrying = RetryOutputParser.from_llm(parser=parser, llm=llm)

# 8) CommaSeparatedListOutputParser — simple list of strings
from langchain_core.output_parsers import CommaSeparatedListOutputParser
parser = CommaSeparatedListOutputParser()
chain  = ChatPromptTemplate.from_template('Five colors:\n{format_instructions}').partial(
    format_instructions=parser.get_format_instructions()
) | llm | parser
chain.invoke({})              # ['red', 'blue', 'green', 'yellow', 'purple']

# 9) Custom parser
from langchain_core.output_parsers import BaseOutputParser

class YesNoParser(BaseOutputParser[bool]):
    def parse(self, text: str) -> bool:
        t = text.strip().lower()
        if t.startswith(('yes', 'true', '1')):  return True
        if t.startswith(('no',  'false', '0')): return False
        raise ValueError(f'Could not parse {text!r}')

chain = prompt | llm | YesNoParser()

# 10) Tool calling — the most reliable structured output
from langchain_core.tools import tool

@tool
def record_recipe(title: str, servings: int, ingredients: List[str], steps: List[str]):
    '''Save a recipe.'''
    return f'saved {title}'

llm_with_tools = llm.bind_tools([record_recipe])
result = llm_with_tools.invoke('Pasta dinner for 4')
if result.tool_calls:
    args = result.tool_calls[0]['args']  # already a dict matching the schema

# 11) Common bugs
#   • Putting format instructions ABOVE the user question → model forgets them
#   • Letting temperature > 0 with strict JSON → invalid JSON randomly
#   • Pydantic parser without format_instructions in prompt → model freelances
#   • Trying to parse markdown code-fence JSON without stripping ``` → fails
#   • Not using with_structured_output when the model supports it → harder to maintain

from langchain_core.output_parsers import StrOutputParser, JsonOutputParser
from pydantic import BaseModel

class User(BaseModel):
    name: str
    age: int

parser = JsonOutputParser(pydantic_object=User)
chain = prompt | llm | parser