LangChain Model I/O: Prompts, Messages and Parsing

Keep prompt construction, provider calls and output validation separate. That makes failures easier to diagnose than one function that builds strings, calls an API and parses a result at once.

#Build messages explicitly

This local example formats a prompt without contacting a model:

pythonCopy
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder

prompt = ChatPromptTemplate.from_messages([
    ("system", "Extract only information present in the supplied text."),
    MessagesPlaceholder("history"),
    ("human", "{text}"),
])
messages = prompt.invoke({"history": [], "text": "Alex uses Python and SQL."})
print(messages.to_messages())

MessagesPlaceholder expects message history, not an arbitrary string. Keep template variables consistent with invocation keys. Use partial() when a value is intentionally fixed, and version the prompt alongside its tests.

#Select an output contract

StrOutputParser produces text. JSON parsing produces data but does not prove its truth. When using Pydantic, validate the parsed result explicitly or use a documented structured-output interface for the selected integration.

pythonCopy
from pydantic import BaseModel, Field

class ExtractedSkills(BaseModel):
    name: str | None = None
    skills: list[str] = Field(default_factory=list)

example = ExtractedSkills.model_validate({"name": "Alex", "skills": ["Python", "SQL"]})
print(example.model_dump())

This is local validation of illustrative data, not a model extraction result. An optional field should remain absent or null when the source does not supply it.

#Test adapters, not just imports

A common chat interface does not guarantee identical provider behavior. Check tool support, content block shape, image support and structured-output constraints. Low temperature can reduce variation when supported; it does not guarantee deterministic or factual output.

Fallbacks must respect the same data-processing policy and permissions. Do not silently send private documents to another provider after a timeout.

Model documentation↗ · Chains