Native JSON Mode

Native JSON Mode

Force LLMs to return guaranteed valid JSON using native provider support.

Provider Support

Provider JSON Mode Support
OpenAI Full support
Google AI Gemini Supported
Mistral AI Supported
Ollama Supported
Anthropic Claude Not supported (use Structured Output)

Quick Start

(require '[langchain4clj.core :as llm])
(require '[clojure.data.json :as json])
(import 'dev.langchain4j.model.chat.request.ResponseFormat)

(def model (llm/create-model {:provider :openai
                              :api-key (System/getenv "OPENAI_API_KEY")}))

(def response (llm/chat model "Return user data as JSON"
                {:response-format ResponseFormat/JSON}))

(def data (json/read-str response :key-fn keyword))
;; => {:name "John Doe" :age 30 :email "john@example.com"}

Usage Methods

;; Method 1: Direct ResponseFormat
(llm/chat model "List 3 programming languages"
  {:response-format ResponseFormat/JSON})

;; Method 2: with-json-mode helper
(llm/chat model "Return user data"
  (llm/with-json-mode {:temperature 0.7}))

;; Method 3: Threading style
(-> {:temperature 0.7}
    llm/with-json-mode
    (as-> opts (llm/chat model "Return data" opts)))

Parsing Responses

;; Using clojure.data.json
(require '[clojure.data.json :as json])
(def data (json/read-str response :key-fn keyword))

;; Using cheshire
(require '[cheshire.core :as json])
(def data (json/parse-string response true))

;; With error handling
(try
  (json/read-str response :key-fn keyword)
  (catch Exception e
    (println "JSON parsing failed:" (.getMessage e))
    nil))

Prompt Engineering

;; Specify the schema in prompt
(llm/chat model 
  "Return JSON with fields: name (string), age (integer), active (boolean)"
  {:response-format ResponseFormat/JSON})

;; Request arrays
(llm/chat model
  "Return JSON array of 3 programming languages with name and year fields"
  {:response-format ResponseFormat/JSON})

;; Nested structures
(llm/chat model
  "Return JSON with user object containing: name, contact (with email and phone)"
  {:response-format ResponseFormat/JSON})

Common Use Cases

Configuration Generation

(def config-json
  (llm/chat model
    "Generate a database configuration with host, port, database name"
    {:response-format ResponseFormat/JSON}))

(def config (json/read-str config-json :key-fn keyword))

Data Transformation

(defn csv-to-json [csv-data]
  (llm/chat model
    (str "Convert this CSV to JSON: " csv-data)
    {:response-format ResponseFormat/JSON}))

Combining with Other Features

;; JSON Mode + Temperature
(llm/chat model "Generate creative product names"
  {:response-format ResponseFormat/JSON
   :temperature 1.0})

;; JSON Mode + System Messages
(llm/chat model "Return user data"
  {:response-format ResponseFormat/JSON
   :system-message "You are a data generator. Always use realistic values."})

When to Use

Use JSON Mode when:

  • You need guaranteed valid JSON
  • Working with supported providers (OpenAI, Gemini, Ollama)
  • Simple JSON structures

Use Structured Output instead when:

  • Need schema validation
  • Provider doesn’t support JSON mode (Anthropic)
  • Need retry logic or EDN output