Skip to content

Latest commit

 

History

History
373 lines (284 loc) · 15.9 KB

README.md

File metadata and controls

373 lines (284 loc) · 15.9 KB

@vscode/prompt-tsx

This library enables you to declare prompts using TSX when you develop VS Code extensions that integrate with Copilot Chat. To learn more, check out our documentation or fork our quickstart sample.

Why TSX?

As AI engineers, our products communicate with large language models using chat messages composed of text prompts. While developing Copilot Chat, we've found that composing prompts with just bare strings is unwieldy and frustrating.

Some of the challenges we ran into include:

  1. We used either programmatic string concatenation or template strings for composing prompts. Programmatic string concatenation made prompt text increasingly difficult to read, maintain, and update over time. Template string-based prompts were rigid and prone to issues like unnecessary whitespace.
  2. In both cases, our prompts and RAG-generated context could not adapt to changing context window constraints as we upgraded our models. Prompts are ultimately bare strings, which makes them hard to edit once they are composed via string concatenation.

To improve the developer experience for writing prompts in language model-based VS Code extensions like Copilot Chat, we built the TSX-based prompt renderer that we've extracted in this library. This has enabled us to compose expressive, flexible prompts that cleanly convert to chat messages. Our prompts are now able to evolve with our product and dynamically adapt to each model's context window.

Key concepts

In this library, prompts are represented as a tree of TSX components that are flattened into a list of chat messages. Each TSX node in the tree has a priority that is conceptually similar to a zIndex (higher number == higher priority).

If a rendered prompt has more message tokens than can fit into the available context window, the prompt renderer prunes messages with the lowest priority from the ChatMessages result, preserving the order in which they were declared. This means your extension code can safely declare TSX components for potentially large pieces of context like conversation history and codebase context.

TSX components at the root level must render to ChatMessages at the root level. ChatMessages may have TSX components as children, but they must ultimately render to text. You can also have TextChunks within ChatMessages, which allows you to reduce less important parts of a chat message under context window limits without losing the full message.

Usage

Workspace Setup

You can install this library in your extension using the command

npm install --save @vscode/prompt-tsx

This library exports a renderPrompt utility for rendering a TSX component to vscode.LanguageModelChatMessages.

To enable TSX use in your extension, add the following configuration options to your tsconfig.json:

{
	"compilerOptions": {
		// ...
		"jsx": "react",
		"jsxFactory": "vscpp",
		"jsxFragmentFactory": "vscppf"
	}
	// ...
}

Note: if your codebase depends on both @vscode/prompt-tsx and another library that uses JSX, for example in a monorepo where a parent folder has dependencies on React, you may encounter compilation errors when trying to add this library to your project. This is because by default, TypeScript includes all @types packages during compilation. You can address this by explicitly listing the types that you want considered during compilation, e.g.:

{
	"compilerOptions": {
		"types": ["node", "jest", "express"]
	}
}

Rendering a Prompt

Next, your extension can use renderPrompt to render a TSX prompt. Here is an example of using TSX prompts in a Copilot chat participant that suggests SQL queries based on database context:

import { renderPrompt } from '@vscode/prompt-tsx';
import * as vscode from 'vscode';
import { TestPrompt } from './prompt';

const participant = vscode.chat.createChatParticipant(
	'mssql',
	async (
		request: vscode.ChatRequest,
		context: vscode.ChatContext,
		response: vscode.ChatResponseStream,
		token: vscode.CancellationToken
	) => {
		response.progress('Reading database context...');

		const models = await vscode.lm.selectChatModels({ family: 'gpt-4' });
		if (models.length === 0) {
			// No models available, return early
			return;
		}
		const chatModel = models[0];

		// Render TSX prompt
		const { messages } = await renderPrompt(
			TestPrompt,
			{ userQuery: request.prompt },
			{ modelMaxPromptTokens: 4096 },
			chatModel
		);

		const chatRequest = await chatModel.sendChatRequest(messages, {}, token);

		// ... Report stream data to VS Code UI
	}
);

Here is how you would declare the TSX prompt rendered above:

import {
	AssistantMessage,
	BasePromptElementProps,
	PromptElement,
	PromptSizing,
	UserMessage,
} from '@vscode/prompt-tsx';
import * as vscode from 'vscode';

export interface PromptProps extends BasePromptElementProps {
	userQuery: string;
}

export interface PromptState {
	creationScript: string;
}

export class TestPrompt extends PromptElement<PromptProps, PromptState> {
	override async prepare() {}

	async render(state: PromptState, sizing: PromptSizing) {
		const sqlExtensionApi = await vscode.extensions.getExtension('ms-mssql.mssql')?.activate();
		const creationScript = await sqlExtensionApi.getDatabaseCreateScript?.();

		return (
			<>
				<AssistantMessage>
					You are a SQL expert.
					<br />
					Your task is to help the user craft SQL queries that perform their task.
					<br />
					You should suggest SQL queries that are performant and correct.
					<br />
					Return your suggested SQL query in a Markdown code block that begins with ```sql and ends
					with ```.
					<br />
				</AssistantMessage>
				<UserMessage>
					Here are the creation scripts that were used to create the tables in my database. Pay
					close attention to the tables and columns that are available in my database:
					<br />
					{state.creationScript}
					<br />
					{this.props.userQuery}
				</UserMessage>
			</>
		);
	}
}

Please note:

  • If your prompt does asynchronous work e.g. VS Code extension API calls or additional requests to the Copilot API for chunk reranking, you can precompute this state in an optional async prepare method. prepare is called before render and the prepared state will be passed back to your prompt component's sync render method.
  • Newlines are not preserved in JSX text or between JSX elements when rendered, and must be explicitly declared with the builtin <br /> attribute.

Prioritization

If a rendered prompt has more message tokens than can fit into the available context window, the prompt renderer prunes messages with the lowest priority from the ChatMessages result.

In the above example, each message had the same priority, so they would be pruned in the order in which they were declared, but we could control that by passing a priority to element:

<>
	<AssistantMessage priority={300}>You are a SQL expert...</AssistantMessage>
	<UserMessage priority={200}>
		Here are the creation scripts that were used to create the tables in my database...
	</UserMessage>
	<UserMessage priority={100}>{this.props.userQuery}</UserMessage>
</>

In this case, a very long userQuery would get pruned from the output first if it's too long. Priorities are local in the element tree, so for example the tree of nodes...

<UserMessage priority={1}>
  <TextChunk priority={100}>A</TextChunk>
  <TextChunk priority={0}>B</TextChunk>
</UserMesssage>
<SystemMessage priority={2}>
  <TextChunk priority={200}>C</TextChunk>
  <TextChunk priority={20}>D</TextChunk>
</SystemMessage>

...would be pruned in the order B->A->D->C. If two sibling elements share the same priority, the renderer looks ahead at their direct children and picks whichever one has a child with the lowest priority: if the SystemMessage and UserMessage in the above example did not declare priorities, the pruning order would be B->D->A->C.

Continuous text strings and elements can both be pruned from the tree. If you have a set of elements that you want to either be include all the time or none of the time, you can use the simple Chunk utility element:

<Chunk>
  The file I'm editing is: <FileLink file={f}>
</Chunk>

Passing Priority

In some cases, you may have logical wrapper elements which container other elements which should share the parent's priority scope. You can use the passPriority attribute for this:

class MyContainer extends PromptElement {
	render() {
		return <>{this.props.children}</>;
	}
}

const myPrompt = (
	<UserMessage>
		<MyContainer passPriority>
			<ChildA priority={1} />
			<ChildB priority={3} />
		</MyContainer>
		<ChildC priority={2} />
	</UserMessage>
);

In this case where we have a wrapper element which includes the children in its own output, the prune order would be ChildA, ChildC, then ChildB.

Flex Behavior

Wholesale pruning is not always ideal. Instead, we'd prefer to include as much of the query as possible. To do this, we can use the flexGrow property, which allows an element to use the remainder of its parent's token budget when it's rendered.

prompt-tsx provides a utility component that supports this use case: TextChunk. Given input text, and optionally a delimiting string or regular expression, it'll include as much of the text as possible to fit within its budget:

<>
	<AssistantMessage priority={300}>You are a SQL expert...</AssistantMessage>
	<UserMessage priority={200}>
		Here are the creation scripts that were used to create the tables in my database...
	</UserMessage>
	<UserMessage priority={100}>
		<TextChunk breakOn=" ">{this.props.userQuery}</TextChunk>
	</UserMessage>
</>

When flexGrow is set for an element, other elements are rendered first, and then the flexGrow element is rendered and given the remaining unused token budget from its container as a parameter in the PromptSizing passed to its prepare and render methods. Here's a simplified version of the TextChunk component:

class SimpleTextChunk extends PromptElement<{ text: string }, string> {
	prepare(sizing: PromptSizing): Promise<string> {
		const words = text.split(' ');
		let str = '';

		for (const word of words) {
			if (tokenizer.tokenLength(str + ' ' + word) > sizing.tokenBudget) {
				break;
			}

			str += ' ' + word;
		}

		return str;
	}

	render(content: string) {
		return <>{content}</>;
	}
}

There are a few similar properties which control budget allocation you mind find useful for more advanced cases:

  • flexReserve: controls the number of tokens reserved from the container's budget before this element gets rendered. For example, if you have a 100 token budget and the elements <><Foo /><Bar flexGrow={1} flexReserve={30} /></>, then Foo would receive a PromptSizing.tokenBudget of 70, and Bar would receive however many tokens of the 100 that Foo didn't use. This is only useful in conjunction with flexGrow.

    This may also be set to a string in the form /N to take a proportion of the container's budget. For example, <Bar flexReserve='/3' flexGrow={1} /> would reserve a third of the container's budget for this element.

  • flexBasis: controls the proportion of tokens allocated from the container's budget to this element. It defaults to 1 on all elements. For example, if you have the elements <><Foo /><Bar /></> and a 100 token budget, each element would be allocated 50 tokens in its PromptSizing.tokenBudget. If you instead render <><Foo /><Bar flexBasis={2} /></>, Bar would receive 66 tokens and Foo would receive 33.

It's important to note that all of the flex* properties allow for cooperative use of the token budget for a prompt, but have no effect on the prioritization and pruning logic undertaken once all elements are rendered.

Local Priority Limits

prompt-tsx provides a TokenLimit element that can be used to set a hard cap on the number of tokens that can be consumed by a prompt or part of a prompt. Using it is fairly straightforward:

class PromptWithLimit extends PromptElement {
	render() {
		return (
			<UserMessage>
				<TokenLimit max={1000}>{/* Your elements here! */}</TokenLimit>
			</UserMessage>
		);
	}
}

TokenLimit subtrees are pruned before the prompt gets pruned. As you would expect, the PromptSizing of child elements inside of a limit reflect the reduced budget. If the TokenLimit would get tokenBudget smaller than its maximum via the usual distribution rules, then that's given it child elements instead (but pruning to the max value still happens.)

Expandable Text

The tools provided by flex* attributes are good, but sometimes you may still end up with unused space in your token budget that you'd like to utilize. We provide a special <Expandable /> element that can be used in this case. It takes a callback that can return a text string.

<Expandable value={async sizing => {
  let data = 'hi';
  while (true) {
    const more = getMoreUsefulData();
    if (await sizing.countTokens(data + more) > sizing.tokenBudget) { break }
    data += more;
  }
  }
  return data;
}} />

After the prompt is rendered, the renderer sums up the tokens used by all messages. If there is unused budget, then any <Expandable /> elements' values are called again with their PromptSizing is increased by the token excess.

If there are multiple <Expandable /> elements, then they're re-called in the order in which they were initially rendered. Because they're designed to fill up any remaining space, it usually makes sense to have at most one <Expandable /> element per prompt.

Debugging Budgeting

You can set a tracer property on the PromptElement to debug how your elements are rendered and how this library allocates your budget. We include a basic HTMLTracer you can use, which can be served on an address:

const renderer = new PromptRenderer(/* ... */);
const tracer = new HTMLTracer();
renderer.tracer = tracer;
renderer.render(/* ... */);

tracer.serveHTML().then(server => {
	console.log('Server address:', server.address);
});

Usage in Tools

Visual Studio Code's API supports language models tools, sometimes called 'functions'. The tools API allows tools to return multiple content types of data to its consumers, and this library supports both returning rich prompt elements to tool callers, as well as using rich content returned from tools.

As a Tool

As a tool, you can use this library normally. However, to return data to the tool caller, you will want to use a special function renderElementJSON to serialize your elements to a plain, transferrable JSON object that can be used by a consumer if they also leverage prompt-tsx:

Note that when VS Code invokes your language model tool, the options may contain tokenOptions which you should pass through as the third argument to renderElementJSON:

// 1. Import prompt-tsx's well-known content type:
import { contentType } from '@vscode/prompt-tsx';

async function doToolInvocation(
	options: LanguageModelToolInvocationOptions
): vscode.LanguageModelToolResult {
	return {
		// In constructing your response, render the tree as JSON.
		[contentType]: await renderElementJSON(MyElement, options.parameters, options.tokenOptions),
		toString: () => '...',
	};
}

As a Consumer

You may invoke the vscode.lm.invokeTool API however you see fit. If you know your token budget in advance, you should pass it to the tool when you call invokeTool via the tokenOptions option. You can then render the result using the <ToolResult /> helper element, for example:

class MyElement extends PromptElement {
	async render(_state: void, sizing: PromptSizing) {
		const result = await vscode.lm.invokeTool(toolId, {
			parameters: getToolParameters(),
			tokenOptions: {
				tokenBudget: sizing.tokenBudget,
				countTokens: (text, token) => sizing.countTokens(text, token),
			},
		});

		return <ToolResult data={result} priority={20} />;
	}
}