How TypeChat Turns Natural Language into Structured JSON with LLMs

Introducing TypeChat

TypeChat is an open‑source library that enables developers to obtain structured responses from large language models (LLMs) and use them as natural‑language interfaces for applications.

Why Structured Natural‑Language Interfaces Matter

Typical UI flows require users to select options step‑by‑step (e.g., coffee type, temperature, sugar level). A free‑form request such as “Give me a large iced latte with half sugar” is ambiguous for a system that only understands structured data.

The order system expects a JSON payload like:

{
  "type": "LatteDrinks",
  "name": "latte",
  "temperature": "iced",
  "quantity": "1",
  "options": { ... }
}

TypeChat’s role is to translate the natural‑language sentence into this exact structure using an LLM.

Traditional AI vs. LLM‑Based Approach

Conventional pipelines rely on manually defined intents, slot‑filling, and large labeled datasets. The workflow typically involves:

Defining intents (e.g., “book ticket”) and associated slots.

Collecting and annotating thousands of example utterances.

Training a classifier to map user input to intents and slots.

This process is labor‑intensive and often struggles with semantic nuances, especially in languages with high ambiguity.

LLM‑driven methods like TypeChat provide two major advantages:

Reduced complexity: The LLM handles semantic understanding and data extraction, eliminating the need for extensive labeling.

Superior semantic grasp: Large models can interpret ambiguous or context‑rich language with far higher accuracy.

How TypeChat Works

Developers first define TypeScript types that represent the structured data their application expects. TypeChat then executes a three‑step pipeline:

Prompt construction: It builds a prompt that embeds the user‑defined types and asks the LLM to produce a matching response.

LLM response & validation: The model’s output is checked against the type definition; if validation fails, additional LLM interactions refine the result.

Result aggregation: Once the output conforms to the type, the structured JSON is returned to the application.

The validated JSON can be fed directly into business logic.

Practical Example

For the coffee‑ordering scenario, a user says “Give me a large iced latte with half sugar.” TypeChat prompts the LLM, receives a JSON payload matching the LatteDrinks type, validates it, and returns the structured data for further processing.

Getting Started

Clone the repository and follow the provided examples. The library works with any LLM accessible via an API (e.g., OpenAI’s ChatGPT).

Repository:

https://github.com/microsoft/TypeChat