JSON Block
Allow LLMs to reply with JSON, which can be rendered as custom components in your application.
Buttons: γ{ type: "buttons", buttons:[ {text:"Star β"}, {text:"Confetti π"} ] }γ
5x
Installation
pnpm add @llm-ui/json
Quick start
Install dependencies
pnpm add @llm-ui/json @llm-ui/react @llm-ui/markdown react-markdown remark-gfm html-react-parser zod
Step 1: Create a markdown component
Create a component to render markdown using react-markdown
.
import ReactMarkdown from "react-markdown";
import remarkGfm from "remark-gfm";
import { type LLMOutputComponent } from "@llm-ui/react";
// Customize this component with your own styling
const MarkdownComponent: LLMOutputComponent = ({ blockMatch }) => {
const markdown = blockMatch.output;
return <ReactMarkdown remarkPlugins={[remarkGfm]}>{markdown}</ReactMarkdown>;
};
Read more in the markdown block docs
Step 2: Setup your custom blockβs JSON schema
Use zod to create a schema for your custom block.
Weβll set up a βbuttonsβ block:
// buttonsSchema.ts
import z from "zod";
const buttonsSchema = z.object({
type: z.literal("buttons"),
buttons: z.array(z.object({ text: z.string() })),
});
Example JSON for your block:
{
"type": "buttons",
"buttons": [{ "text": "Button 1" }, { "text": "Button 2" }]
}
Step 3: Create a custom block component
import { parseJson5 } from "@llm-ui/json";
import { type LLMOutputComponent } from "@llm-ui/react";
// Customize this component with your own styling
const ButtonsComponent: LLMOutputComponent = ({ blockMatch }) => {
if (!blockMatch.isVisible) {
return null;
}
// use buttonsSchema from step 2
const { data: buttons, error } = buttonsSchema.safeParse(
parseJson5(blockMatch.output),
);
if (error) {
return <div>{error.toString()}</div>;
}
return (
<div>
{buttons.buttons.map((button, index) => (
<button key={index}>{button.text}</button>
))}
</div>
);
};
Step 4: Render custom blocks with llm-ui
Now weβve created our components, weβre ready to use useLLMOutput to render language model output which contains markdown and buttons components.
import { markdownLookBack } from "@llm-ui/markdown";
import { useLLMOutput, useStreamExample } from "@llm-ui/react";
import { jsonBlock } from "@llm-ui/json";
const example = `Buttons:
γ{type:"buttons",buttons:[{text:"Star β"}, {text:"Confetti π"}]}γ
`;
const Example = () => {
const { isStreamFinished, output } = useStreamExample(example);
const { blockMatches } = useLLMOutput({
llmOutput: output,
blocks: [
{
...jsonBlock({type: "buttons"}),
component: ButtonsComponent, // from step 3
},
],
fallbackBlock: {
lookBack: markdownLookBack(),
component: MarkdownComponent, // from step 1
},
isStreamFinished,
});
return (
<div>
{blockMatches.map((blockMatch, index) => {
const Component = blockMatch.block.component;
return <Component key={index} blockMatch={blockMatch} />;
})}
</div>
);
}
Step 5: Prompt LLM with your custom block
Generate the prompt for your JSON block:
import { jsonBlockPrompt } from "@llm-ui/json";
const prompt = jsonBlockPrompt({
name: "Button",
schema: buttonsSchema,
examples: [
{ type: "buttons", buttons: [{ text: "Button 1" }, { text: "Button 2" }] },
],
options: {
type: "buttons",
},
});
Generates:
You can respond with a Button component by wrapping JSON in γγ.
The JSON schema is:
{"type":"object","properties":{"type":{"type":"string","const":"buttons"},"buttons":{"type":"array","items":{"type":"object","properties":{"text":{"type":"string"}},"required":["text"],"additionalProperties":false}}},"required":["type","buttons"]}
Examples:
γ{"type":"buttons","buttons":[{"text":"Button 1"},{"text":"Button 2"}]}γ
You can also hardcode the prompt into your
Options
{
// Required
type: "buttons",
// Optional, defaults:
startChar: "γ",
endChar: "γ",
typeKey: "type", // the key in the JSON object which determines the block type e.g. {"type": "buttons"}
defaultVisible: false, // See below
visibleKeyPaths: [], // See below
invisibleKeyPaths: [], // See below
}
defaultVisible
defaultVisible: false
Generate no βvisibleTextβ until the whole block is parsed.
When a partial block is parsed:
γ{ type:"buttons", buttons: [{text:"Button 1"}, {text:"But `
blockMatch.visibleText;
// => ""
blockMatch.isVisible;
// => false
// llm-ui completes the partial JSON
blockMatch.output;
// => "{ type:"buttons", buttons: [{text:"Button 1"}, {text:"But"}] }"
When the whole block is parsed:
γ{ type:"buttons", buttons: [{text:"Button 1"}, {text:"Button 2"}] }γ
blockMatch.visibleText;
// => " "
blockMatch.isVisible;
// => true
blockMatch.output;
// => "{ type:"buttons", buttons: [{text:"Button 1"}, {text:"Button 2"}] }"
defaultVisible: true
Generates βvisibleTextβ as the reponse is parsed:
γ{ type:"buttons", buttons: [{text:"Button 1"}, {text:"But
blockMatch.visibleText;
// => "B"
// then
// => "Bu"
// then
// => "But"
blockMatch.isVisible;
// => true
blockMatch.output;
// => "{ type:"buttons", buttons: [{text:"B"}] }"
// then
// => "{ type:"buttons", buttons: [{text:"Bu"}] }"
// then
// => "{ type:"buttons", buttons: [{text:"But"}] }"
visibleKeyPaths
and invisibleKeyPaths
You can use visibleKeyPaths
and invisibleKeyPaths
to determine which fields in the JSON object are visible or invisible.
The path syntax is jsonpath.
{
defaultVisible: false,
visibleKeyPaths: ["$.buttons[*].text"],
}
{
defaultVisible: true, // typeKey e.g. "type" is always invisible
invisibleKeyPaths: ["$.buttons[*].color"],
}
Prompts
jsonBlockPrompt
Returns a full prompt to send to the LLM.
import { jsonBlockPrompt } from "@llm-ui/json";
import z from "zod";
jsonBlockPrompt({
name: "Button",
schema: z.object({
type: z.literal("buttons"),
buttons: z.array(z.object({ text: z.string() })),
}),
examples: [
{ type: "buttons", buttons: [{ text: "Button 1" }, { text: "Button 2" }] },
],
options: {
type: "buttons",
startChar: "γ",
endChar: "γ",
typeKey: "type",
defaultVisible: false,
},
});
// =>
"
You can respond with a Button component by wrapping JSON in γγ.
The JSON schema is:
{"type":"object","properties":{"type":{"type":"string","const":"buttons"},"buttons":{"type":"array","items":{"type":"object","properties":{"text":{"type":"string"}},"required":["text"],"additionalProperties":false}}},"required":["type","buttons"]}
Examples:
γ{"type":"buttons","buttons":[{"text":"Button 1"},{"text":"Button 2"}]}γ
"
jsonBlockSchema
Returns the schema as a JSON schema string.
import { jsonBlockSchema } from "@llm-ui/json";
import z from "zod";
jsonBlockSchema(z.object({
type: z.literal("buttons"),
buttons: z.array(z.object({ text: z.string() })),
}));
// =>
{"type":"object","properties":{"type":{"type":"string","const":"buttons"},"buttons":{"type":"array","items":{"type":"object","properties":{"text":{"type":"string"}},"required":["text"],"additionalProperties":false}}},"required":["type","buttons"]}
jsonBlockExample
Generates a single JSON block usage example.
import { jsonBlockExample } from "@llm-ui/json";
import z from "zod";
jsonBlockExample({
schema: z.object({
type: z.literal("buttons"),
buttons: z.array(z.object({ text: z.string() })),
}),
example: { type: "buttons", buttons: [{ text: "Button 1" }, { text: "Button 2" }] },
options: {
type: "buttons",
startChar: "γ",
endChar: "γ",
typeKey: "type",
defaultVisible: false,
},
});
// =>
γ{"type":"buttons","buttons":[{"text":"Button 1"},{"text":"Button 2"}]}γ
Json block functions
jsonBlock
Returns a JSON block object to be used by useLLMOutput
.
import { jsonBlock } from "@llm-ui/json";
const options = {
type: "buttons",
startChar: "γ",
endChar: "γ",
typeKey: "type",
defaultVisible: false,
};
jsonBlock(options);
// =>
{
findCompleteMatch: findCompleteJsonBlock(options),
findPartialMatch: findPartialJsonBlock(options),
lookBack: jsonBlockLookBack(options),
component: () => <div>Json block</div>,
}
Accepts options parameter.
findCompleteJsonBlock
Finds a complete JSON block in a string:
findCompleteJsonBlock({ type: "buttons" });
Will match:
γ{"type":"buttons","buttons":[{"text":"Button 1"},{"text":"Button 2"}]}γ
Accepts options parameter.
findPartialJsonBlock
Find a partial JSON block in a string.
findPartialJsonBlock({ type: "buttons" });
Will match:
γ{"type":"buttons","buttons":[{"text":"Button 1"},{"text":
Accepts options parameter.
jsonBlockLookBack
Look back function for the JSON block.
jsonBlockLookBack({ type: "buttons" });
Accepts options parameter.
Parse
parseJson5
Parse a JSON5 output string.
import { parseJson5 } from "@llm-ui/json";
parseJson5(
'{type:"buttons",buttons:[{"text":"Button 1"},{"text":"Button 2"}]}',
);
// =>
{
type: "buttons",
buttons: [{ text: "Button 1" }, { text: "Button 2" }],
}
Optimizing for performance
When using the JSON block youβll want to consider the performance of your application.
JSON has a high overhead of non-content characters, this may result in a noticable pause as llm-ui waits for the full block.
You may wish to use skeleton or other loading states to indicate to the user that the block is being parsed.
In production we also recommend using a JSON format which is as concise as possible to reduce the overhead of non-content characters.
For example:
γ{"type":"buttons","buttons":[{"text":"Button 1"},{"text":...
Becomes:
γ{"t":"b","bs":[{"t":"Button 1"},{"t":...
You could also consider using the CSV block for a lower overhead format.