langchain-pydantic-minifier 0.1.4

automatically reduces token usage by replacing field names with short identifiers (e.g., a, b, c, …), while preserving full descriptions and reversibility.

🚀 Add MinifiedPydanticOutputParser to Optimize Token Usage in LLM Outputs

This PR introduces a new class: MinifiedPydanticOutputParser, a drop-in replacement for PydanticOutputParser that automatically reduces token usage by replacing field names with short identifiers (e.g., a, b, c, …), while preserving full descriptions and reversibility.

✨ What It Does

• Transforms a given Pydantic schema by replacing verbose field names with shorter aliases.
• Retains all Field(..., description=...) information — essential for prompt construction and LLM understanding.
• Accepts minified JSON outputs from the LLM and reconstructs the original schema transparently.
• Supports nested models and list fields recursively.
• Compatible with strict=True mode used with with_structured_output.

✅ Benefits

• Reduces prompt and completion token count, leading to faster LLM response times and lower inference costs.
• Maintains clarity in the LLM's understanding of the field semantics thanks to preserved descriptions.
• No code change required downstream — consumers receive the original schema post-parsing.

📉 Performance Impact

In personal benchmarks, this optimization led to a ~30% reduction in LLM response time, due to:

• Fewer tokens needing generation
• Reduced I/O and parsing overhead

💸 This also translates to lower API costs, especially in high-throughput or large-output scenarios.

Examples

• See Gemini example
• see OpenAI OutputParser example
• see OpenAI structured_output strict example

🧪 What it does

Given:

class User(BaseModel):
    first_name: str = Field(..., description="The user's first name")
    last_name: str = Field(..., description="The user's last name")

The model is transformed into:

class MinifiedUser(BaseModel):
    a: str = Field(..., description="The user's first name")
    b: str = Field(..., description="The user's last name")

Then seamlessly restored to the original User class after parsing the LLM output.