🧭Guidance

NEW

Also known as: Microsoft Guidance

An open-source library to programmatically control LLM generation using templates, constraints, and structured output.

DESCRIPTION

Guidance lets developers interleave prompt text and program logic to constrain model outputs. It supports templating, JSON or schema-constrained generation, and deterministic structures. The library integrates with popular LLM providers and toolchains to enforce format and content rules. It helps build reliable applications requiring structured outputs and guardrails.

ALTERNATIVE DEFINITIONS

Microsoft

Guidance is a Python library for controlling modern language models.

Highlights control of LLMs as the central capability.

Source

Microsoft

Guidance provides a programming model for authoring complex prompting pipelines with constraints and structure.

Emphasizes programming model and constraints for complex prompting pipelines.

Source

RELATED TERMS

Prompt Engineering Structured Output Template JSON Schema Guardrails Function Calling LangChain Retrieval-Augmented Generation

COMPETING TERMS

LMQL Guardrails AI Outlines

RELATIONS

• Conforms to: JSON Schema

KNOWLEDGE AREAS

Platform & Tooling AI MLOps

TARGET AUDIENCE

AI Engineer Software Developer MLOps Engineer

EXTERNAL SOURCES

• Wikipedia - Prompt engineering
• Microsoft - Official repository and README
• Microsoft - Project docs and examples

TOP QUESTIONS

Why use Guidance for LLM applications?

It enforces structure and constraints, reducing parsing errors and hallucinations. This improves reliability for production-grade workflows.

Category: Relevancy

How does Guidance compare to LMQL or Guardrails?

Guidance focuses on template-based control with programmatic constraints. LMQL offers a query language; Guardrails emphasizes validation frameworks.

Category: Comparison

Can Guidance enforce JSON schema outputs?

Yes. It provides structured generation features that align outputs to specified schemas.

Category: Clarification

What integrations does Guidance support?

It works with major LLM providers and Python ecosystems. Check the docs for supported backends.

Category: Tools

What are common mistakes when using Guidance?

Overly rigid constraints can hurt quality; unclear templates cause inconsistent outputs. Benchmarking templates is essential.

Category: Mistakes

How do you test templates built with Guidance?

Use unit tests, golden files, and randomized prompts. Validate structure, correctness, and robustness under varied inputs.

Category: Testing

Is Guidance suitable for RAG pipelines?

Yes. It can constrain answer formats, citations, and reasoning steps, complementing retrievers and readers.

Category: UseCases

How to start learning Guidance?

Read the README and examples, replicate sample pipelines, and gradually add constraints. Evaluate against baseline prompts.

Category: GettingStarted

Does Guidance reduce hallucination risks?

Constraints and structured outputs help, but retrieval, calibration, and safety filters are still necessary.

Category: Risk