Output Parsers
LLMs generate text, but applications need JSON, Pydantic models, lists, or enums. Parsers extract structured data from LLM output, with retry logic when the model doesn't format correctly.
A translator converting a chef's verbal recipe (free text) into a printed recipe card with labeled sections (structured format). The parser is the translator that extracts ingredients, steps, and timing.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
Parse or Retry
Scroll inside the frame to explore · use + / − to zoom up to 200%
Format instructions go into the prompt. Failed JSON gets a second chance with the error.
Key Takeaways
- 1.Output parsers convert raw LLM text responses into structured Python objects — handling the gap between free-form generation and typed application code.
- 2.Parser types: PydanticOutputParser (parse into Pydantic models), JSONOutputParser, StructuredOutputParser (with schema), CommaSeparatedListOutputParser.
- 3.Flow: parser.get_format_instructions() → inject into prompt → LLM generates → parser.parse(response).
- 4.Retry: if parsing fails, re-prompt with error message.
- 5.Modern alternative: native structured outputs (OpenAI response_format, tool calling) are more reliable than parse-from-text.
Real Example
Scenario
Extract meeting info: prompt asks for title, date, attendees. LLM returns JSON string. PydanticOutputParser validates into Meeting(title=str, date=datetime, attendees=list[str]). If date format wrong, parser raises error → retry with 'date must be YYYY-MM-DD'.
What you would do
In LLM Engineering & APIs, apply Output Parsers to this scenario: Extract meeting info: prompt asks for title, date, attendees. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Practice Task
Open the Code Walkthrough below and run it locally. Change one parameter related to Output Parsers (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.
Code Walkthrough
Highlighted lines show where Output Parsers happens in the code.
1from langchain_core.output_parsers import PydanticOutputParser # import dependencies2from pydantic import BaseModel, Field # import dependencies3
4class MovieReview(BaseModel): # define a data structure or component5 title: str = Field(description="Movie title")6 rating: int = Field(description="Rating 1-10")7 summary: str = Field(description="One sentence summary")8
9parser = PydanticOutputParser(pydantic_object=MovieReview) # key line for Output Parsers10
11format_instructions = parser.get_format_instructions()12# Inject into prompt: f"\n{format_instructions}"13
14# Simulate LLM response15llm_output = '{"title": "Inception", "rating": 9, "summary": "Mind-bending thriller about dreams."}' # key line for Output Parsers16review = parser.parse(llm_output) # key line for Output Parsers17print(review.title, review.rating) # show output for debuggingCommands
Commands to Remember
from langchain_core.output_parsers import PydanticOutputParser # text to typed objectparser.get_format_instructions() # inject schema into the promptparser.parse(llm_output) # raises if JSON does not matchPrefer response_format json_schema when the provider supports it # fewer retries
Cheat Sheet
Quick recap
quick ref- •PydanticOutputParser(pydantic_object=Model)
- •parser.get_format_instructions()
- •parser.parse(llm_output)
- •Retry on ValidationError
- •Prefer structured outputs API when available
- •Field(description=) guides LLM
Common Mistakes
- ✕Not including format instructions in prompt — low parse success
- ✕No retry logic — single parse failure breaks pipeline
- ✕Using text parsing when structured outputs API available
- ✕Overly complex schemas — LLM struggles with deep nesting