API reference · Kjarni 0.2.7
Class: Kjarni
kjarni-wasm • Docs
kjarni-wasm / Kjarni
Models running in a Web Worker. Create one with Kjarni.load; every method returns a promise, so the page stays responsive while a model works.
const kjarni = await Kjarni.load({ encoder: "/models/minilm.kjq" });
const [a, b] = await kjarni.encode(["doctor", "physician"]);
console.log(Kjarni.cosine(a, b)); // 0.8598
Methods
chat()
chat(
prompt,options):AsyncGenerator<string,void,unknown>
Generates a reply, yielding text as it arrives.
for await (const piece of kjarni.chat("Explain RAG")) out.textContent += piece;
Parameters
• prompt: string
• options: ChatOptions = {}
Returns
AsyncGenerator<string, void, unknown>
Defined in
classify()
classify(
texts):Promise<ClassifyResult[]>
The best label for each input.
Parameters
• texts: string | string[]
Returns
Promise<ClassifyResult[]>
Defined in
close()
close():
Promise<void>
Releases the models and stops the worker.
Returns
Promise<void>
Defined in
contextSize()
contextSize():
Promise<number>
The chat model's context window, in tokens.
Returns
Promise<number>
Defined in
encode()
encode(
texts,normalize):Promise<Float32Array[]>
One vector per input, in the order given.
Parameters
• texts: string[]
• normalize: boolean = true
Returns
Promise<Float32Array[]>
Defined in
labels()
labels():
Promise<string[]>
Every label this classifier can predict, in index order.
Returns
Promise<string[]>
Defined in
rerank()
rerank(
query,documents,limit):Promise<RankedDocument[]>
Documents reordered by relevance, best first. Indices point into your input.
Parameters
• query: string
• documents: string[]
• limit: number = documents.length
Returns
Promise<RankedDocument[]>
Defined in
similarity()
similarity(
a,b):Promise<number>
Cosine similarity between two texts, from -1 to 1.
Parameters
• a: string
• b: string
Returns
Promise<number>
Defined in
cosine()
staticcosine(a,b):number
Cosine similarity between two vectors.
Parameters
• a: Float32Array
• b: Float32Array
Returns
number
Defined in
load()
staticload(options):Promise<Kjarni>
Starts a worker and loads whichever models you name.
Parameters
• options: LoadOptions
Returns
Promise<Kjarni>
Defined in
Kjarni runs the same engine on every platform: