This proposal is an early design sketch by the Chrome built-in AI team to describe the problem below and solicit feedback on the proposed solution. It has not been approved to ship in Chrome.
Browsers and operating systems are increasingly expected to gain access to a language model. (Example, example, example.) Language models are known for their versatility. With enough creative prompting, they can help accomplish tasks as diverse as:
- Classification, tagging, and keyword extraction of arbitrary text;
- Helping users compose text, such as blog posts, reviews, or biographies;
- Summarizing, e.g. of articles, user reviews, or chat logs;
- Generating titles or headlines from article contents
- Answering questions based on the unstructured contents of a web page
- Translation between languages
- Proofreading
Although the Chrome built-in AI team is exploring purpose-built APIs for some of these use cases (e.g. translation, and perhaps in the future summarization and compose), we are also exploring a general-purpose "prompt API" which allows web developers to prompt a language model directly. This gives web developers access to many more capabilities, at the cost of requiring them to do their own prompt engineering.
Currently, web developers wishing to use language models must either call out to cloud APIs, or bring their own and run them using technologies like WebAssembly and WebGPU. By providing access to the browser or operating system's existing language model, we can provide the following benefits compared to cloud APIs:
- Local processing of sensitive data, e.g. allowing websites to combine AI features with end-to-end encryption.
- Potentially faster results, since there is no server round-trip involved.
- Offline usage.
- Lower API costs for web developers.
- Allowing hybrid approaches, e.g. free users of a website use on-device AI whereas paid users use a more powerful API-based model.
Similarly, compared to bring-your-own-AI approaches, using a built-in language model can save the user's bandwidth, likely benefit from more optimizations, and have a lower barrier to entry for web developers.
Even more so than many other behind-a-flag APIs, the prompt API is an experiment, designed to help us understand web developers' use cases to inform a roadmap of purpose-built APIs. However, we want to publish an explainer to provide documentation and a public discussion place for the experiment while it is ongoing.
Our goals are to:
- Provide web developers a uniform JavaScript API for accessing browser-provided language models.
- Abstract away specific details of the language model in question as much as possible, e.g. tokenization, system messages, or control tokens.
- Guide web developers to gracefully handle failure cases, e.g. no browser-provided model being available.
- Allow a variety of implementation strategies, including on-device or cloud-based models, while keeping these details abstracted from developers.
The following are explicit non-goals:
- We do not intend to force every browser to ship or expose a language model; in particular, not all devices will be capable of storing or running one. It would be conforming to implement this API by always signaling that no language model is available, or to implement this API entirely by using cloud services instead of on-device models.
- We do not intend to provide guarantees of language model quality, stability, or interoperability between browsers. In particular, we cannot guarantee that the models exposed by these APIs are particularly good at any given use case. These are left as quality-of-implementation issues, similar to the shape detection API. (See also a discussion of interop in the W3C "AI & the Web" document.)
The following are potential goals we are not yet certain of:
- Allow web developers to know, or control, whether language model interactions are done on-device or using cloud services. This would allow them to guarantee that any user data they feed into this API does not leave the device, which can be important for privacy purposes. Similarly, we might want to allow developers to request on-device-only language models, in case a browser offers both varieties.
- Allow web developers to know some identifier for the language model in use, separate from the browser version. This would allow them to allowlist or blocklist specific models to maintain a desired level of quality, or restrict certain use cases to a specific model.
Both of these potential goals could pose challenges to interoperability, so we want to investigate more how important such functionality is to developers to find the right tradeoff.
In this example, a single string is used to prompt the API, which is assumed to come from the user. The returned response is from the assistant.
const session = await ai.assistant.create();
// Prompt the model and wait for the whole result to come back.
const result = await session.prompt("Write me a poem.");
console.log(result);
// Prompt the model and stream the result:
const stream = await session.promptStreaming("Write me an extra-long poem.");
for await (const chunk of stream) {
console.log(chunk);
}The assistant can be configured with a special "system prompt" which gives it the context for future interactions:
const session = await ai.assistant.create({
systemPrompt: "Pretend to be an eloquent hamster."
});
console.log(await session.prompt("What is your favorite food?"));The system prompt is special, in that the assistant will not respond to it, and it will be preserved even if the context window otherwise overflows due to too many calls to prompt().
If the system prompt is too large (see below), then the promise will be rejected with a "QuotaExceededError" DOMException.
If developers want to provide examples of the user/assistant interaction, they can use the initialPrompts array. This aligns with the common "chat completions API" format of { role, content } pairs, including a "system" role which can be used instead of the systemPrompt option shown above.
const session = await ai.assistant.create({
initialPrompts: [
{ role: "system", content: "Predict up to 5 emojis as a response to a comment. Output emojis, comma-separated." },
{ role: "user", content: "This is amazing!" },
{ role: "assistant", content: "❤️, ➕" },
{ role: "user", content: "LGTM" },
{ role: "assistant", content: "👍, 🚢" }
]
});
// Clone an existing session for efficiency, instead of recreating one each time.
async function predictEmoji(comment) {
const freshSession = await session.clone();
return await freshSession.prompt(comment);
}
const result1 = await predictEmoji("Back to the drawing board");
const result2 = await predictEmoji("This code is so good you should get promoted");Some details on error cases:
- Using both
systemPromptand a{ role: "system" }prompt ininitialPrompts, or using multiple{ role: "system" }prompts, or placing the{ role: "system" }prompt anywhere besides at the 0th position ininitialPrompts, will reject with aTypeError. - If the combined token length of all the initial prompts (including the separate
systemPrompt, if provided) is too large, then the promise will be rejected with a"QuotaExceededError"DOMException.
In addition to the systemPrompt and initialPrompts options shown above, the currently-configurable options are temperature and top-K. More information about the values for these parameters can be found using the capabilities() API explained below.
const customSession = await ai.assistant.create({
temperature: 0.8,
topK: 10
});
const capabilities = await ai.assistant.capabilities();
const slightlyHighTemperatureSession = await ai.assistant.create({
temperature: Math.max(capabilities.defaultTemperature * 1.2, 1.0),
});
// capabilities also contains defaultTopK and maxTopK.Each assistant session consists of a persistent series of interactions with the model:
const session = await ai.assistant.create({
systemPrompt: "You are a friendly, helpful assistant specialized in clothing choices."
});
const result = await session.prompt(`
What should I wear today? It's sunny and I'm unsure between a t-shirt and a polo.
`);
console.log(result);
const result2 = await session.prompt(`
That sounds great, but oh no, it's actually going to rain! New advice??
`);Multiple unrelated continuations of the same prompt can be set up by creating a session and then cloning it:
const session = await ai.assistant.create({
systemPrompt: "You are a friendly, helpful assistant specialized in clothing choices."
});
const session2 = await session.clone();An assistant session can be destroyed, either by using an AbortSignal passed to the create() method call:
const controller = new AbortController();
stopButton.onclick = () => controller.abort();
const session = await ai.assistant.create({ signal: controller.signal });or by calling destroy() on the session:
stopButton.onclick = () => session.destroy();Destroying a session will have the following effects:
-
If done before the promise returned by
create()is settled:-
Stop signaling any ongoing download progress for the language model. (The browser may also abort the download, or may continue it. Either way, no further
downloadprogressevents will fire.) -
Reject the
create()promise.
-
-
Otherwise:
-
Reject any ongoing calls to
prompt(). -
Error any
ReadableStreams returned bypromptStreaming().
-
-
Most importantly, destroying the session allows the user agent to unload the language model from memory, if no other APIs or sessions are using it.
In all cases the exception used for rejecting promises or erroring ReadableStreams will be an "AbortError" DOMException, or the given abort reason.
The ability to manually destroy a session allows applications to free up memory without waiting for garbage collection, which can be useful since language models can be quite large.
Specific calls to prompt() or promptStreaming() can be aborted by passing an AbortSignal to them:
const controller = new AbortController();
stopButton.onclick = () => controller.abort();
const result = await session.prompt("Write me a poem", { signal: controller.signal });Note that because sessions are stateful, and prompts can be queued, aborting a specific prompt is slightly complicated:
- If the prompt is still queued behind other prompts in the session, then it will be removed from the queue.
- If the prompt is being currently processed by the model, then it will be aborted, and the prompt/response pair will be removed from the conversation history.
- If the prompt has already been fully processed by the model, then attempting to abort the prompt will do nothing.
A given assistant session will have a maximum number of tokens it can process. Developers can check their current usage and progress toward that limit by using the following properties on the session object:
console.log(`${session.tokensSoFar}/${session.maxTokens} (${session.tokensLeft} left)`);To know how many tokens a string will consume, without actually processing it, developers can use the countPromptTokens() method:
const numTokens = await session.countPromptTokens(promptString);Some notes on this API:
- We do not expose the actual tokenization to developers since that would make it too easy to depend on model-specific details.
- Implementations must include in their count any control tokens that will be necessary to process the prompt, e.g. ones indicating the start or end of the input.
- The counting process can be aborted by passing an
AbortSignal, i.e.session.countPromptTokens(promptString, { signal }).
It's possible to send a prompt that causes the context window to overflow. That is, consider a case where session.countPromptTokens(promptString) > session.tokensLeft before calling session.prompt(promptString), and then the web developer calls session.prompt(promptString) anyway. In such cases, the initial portions of the conversation with the assistant will be removed, one prompt/response pair at a time, until enough tokens are available to process the new prompt. The exception is the system prompt, which is never removed. If it's not possible to remove enough tokens from the conversation history to process the new prompt, then the prompt() or promptStreaming() call will fail with an "QuotaExceededError" DOMException and nothing will be removed.
Such overflows can be detected by listening for the "contextoverflow" event on the session:
session.addEventListener("contextoverflow", () => {
console.log("Context overflow!");
});In all our above examples, we call ai.assistant.create() and assume it will always succeed.
However, sometimes a language model needs to be downloaded before the API can be used. In such cases, immediately calling create() will start the download, which might take a long time. The capabilities API gives you insight into the download status of the model:
const capabilities = await ai.assistant.capabilities();
console.log(capabilities.available);The capabilities.available property is a string that can take one of three values:
"no", indicating the device or browser does not support prompting a language model at all."after-download", indicating the device or browser supports prompting a language model, but it needs to be downloaded before it can be used."readily", indicating the device or browser supports prompting a language model and it’s ready to be used without any downloading steps.
In the "after-download" case, developers might want to have users confirm before you call create() to start the download, since doing so uses up significant bandwidth and users might not be willing to wait for a large download before using the site or feature.
Note that regardless of the return value of available, create() might also fail, if either the download fails or the session creation fails.
The capabilities API also contains other information about the model:
defaultTemperature,defaultTopK, andmaxTopKproperties giving information about the model's sampling parameters.supportsLanguage(languageTag), which returns"no","after-download", or"readily"to indicate whether the model supports conversing in a given human language.
In cases where the model needs to be downloaded as part of creation, you can monitor the download progress (e.g. in order to show your users a progress bar) using code such as the following:
const session = await ai.assistant.create({
monitor(m) {
m.addEventListener("downloadprogress", e => {
console.log(`Downloaded ${e.loaded} of ${e.total} bytes.`);
});
}
});If the download fails, then downloadprogress events will stop being emitted, and the promise returned by create() will be rejected with a "NetworkError" DOMException.
What's up with this pattern?
This pattern is a little involved. Several alternatives have been considered. However, asking around the web standards community it seemed like this one was best, as it allows using standard event handlers and ProgressEvents, and also ensures that once the promise is settled, the assistant object is completely ready to use.
It is also nicely future-extensible by adding more events and properties to the m object.
Finally, note that there is a sort of precedent in the (never-shipped) FetchObserver design.
// Shared self.ai APIs
partial interface WindowOrWorkerGlobalScope {
[Replaceable] readonly attribute AI ai;
};
[Exposed=(Window,Worker)]
interface AI {
readonly attribute AIAssistantFactory assistant;
};
[Exposed=(Window,Worker)]
interface AICreateMonitor : EventTarget {
attribute EventHandler ondownloadprogress;
// Might get more stuff in the future, e.g. for
// https://github.com/explainers-by-googlers/prompt-api/issues/4
};
callback AICreateMonitorCallback = undefined (AICreateMonitor monitor);
enum AICapabilityAvailability { "readily", "after-download", "no" };// Assistant
[Exposed=(Window,Worker)]
interface AIAssistantFactory {
Promise<AIAssistant> create(optional AIAssistantCreateOptions options = {});
Promise<AIAssistantCapabilities> capabilities();
};
[Exposed=(Window,Worker)]
interface AIAssistant : EventTarget {
Promise<DOMString> prompt(DOMString input, optional AIAssistantPromptOptions options = {});
ReadableStream promptStreaming(DOMString input, optional AIAssistantPromptOptions options = {});
Promise<unsigned long long> countPromptTokens(DOMString input, optional AIAssistantPromptOptions options = {});
readonly attribute unsigned long long maxTokens;
readonly attribute unsigned long long tokensSoFar;
readonly attribute unsigned long long tokensLeft;
readonly attribute unsigned long topK;
readonly attribute float temperature;
attribute EventHandler oncontextoverflow;
Promise<AIAssistant> clone();
undefined destroy();
};
[Exposed=(Window,Worker)]
interface AIAssistantCapabilities {
readonly attribute AICapabilityAvailability available;
// Always null if available === "no"
readonly attribute unsigned long? defaultTopK;
readonly attribute unsigned long? maxTopK;
readonly attribute float? defaultTemperature;
AICapabilityAvailability supportsLanguage(DOMString languageTag);
};
dictionary AIAssistantCreateOptions {
AbortSignal signal;
AICreateMonitorCallback monitor;
DOMString systemPrompt;
sequence<AIAssistantPrompt> initialPrompts;
[EnforceRange] unsigned long topK;
float temperature;
};
dictionary AIAssistantPrompt {
AIAssistantPromptRole role;
DOMString content;
};
dictionary AIAssistantPromptOptions {
AbortSignal signal;
};
enum AIAssistantPromptRole { "system", "user", "assistant" };We intend for this API to expose instruction-tuned models. Although we cannot mandate any particular level of quality or instruction-following capability, we think setting this base expectation can help ensure that what browsers ship is aligned with what web developers expect.
To illustrate the difference and how it impacts web developer expectations:
- In a base model, a prompt like "Write a poem about trees." might get completed with "... Write about the animal you would like to be. Write about a conflict between a brother and a sister." (etc.) It is directly completing plausible next tokens in the text sequence.
- Whereas, in an instruction-tuned model, the model will generally follow instructions like "Write a poem about trees.", and respond with a poem about trees.
To ensure the API can be used by web developers across multiple implementations, all browsers should be sure their models behave like instruction-tuned models.
To actually get a response back from the model given a prompt, the following possible stages are involved:
- Download the model, if necessary.
- Establish a session, including configuring per-session options.
- Add an initial prompt to establish context. (This will not generate a response.)
- Execute a prompt and receive a response.
We've chosen to manifest these 3-4 stages into the API as two methods, ai.assistant.create() and session.prompt()/session.promptStreaming(), with some additional facilities for dealing with the fact that ai.assistant.create() can include a download step. Some APIs simplify this into a single method, and some split it up into three (usually not four).
Our design here uses sessions. An alternate design, seen in some APIs, is to require the developer to feed in the entire conversation history to the model each time, keeping track of the results.
This can be slightly more flexible; for example, it allows manually correcting the model's responses before feeding them back into the context window.
However, our understanding is that the session-based model can be more efficiently implemented, at least for browsers with on-device models. (Implementing it for a cloud-based model would likely be more work.) And, developers can always achieve a stateless model by using a new session for each interaction.
If cloud-based language models are exposed through this API, then there are potential privacy issues with exposing user or website data to the relevant cloud and model providers. This is not a concern specific to this API, as websites can already choose to expose user or website data to other origins using APIs such as fetch(). However, it's worth keeping in mind, and in particular as discussed in our Goals, perhaps we should make it easier for web developers to know whether a cloud-based model is in use, or which one.
If on-device language models are updated separately from browser and operating system versions, this API could enhance the web's fingerprinting service by providing extra identifying bits. Mandating that older browser versions not receive updates or be able to download models from too far into the future might be a possible remediation for this.
Finally, we intend to prohibit (in the specification) any use of user-specific information that is not directly supplied through the API. For example, it would not be permissible to fine-tune the language model based on information the user has entered into the browser in the past.