smartai/ts/provider.openai.ts

196 lines
5.8 KiB
TypeScript
Raw Normal View History

2024-03-31 01:32:37 +01:00
import * as plugins from './plugins.js';
import * as paths from './paths.js';
2024-04-04 02:47:44 +02:00
import { MultiModalModel } from './abstract.classes.multimodal.js';
2024-03-31 01:32:37 +01:00
2024-04-27 12:47:49 +02:00
export interface IOpenaiProviderOptions {
openaiToken: string;
}
2024-04-04 02:47:44 +02:00
export class OpenAiProvider extends MultiModalModel {
2024-04-27 12:47:49 +02:00
private options: IOpenaiProviderOptions;
2024-03-31 01:32:37 +01:00
public openAiApiClient: plugins.openai.default;
2024-04-27 12:47:49 +02:00
public smartpdfInstance: plugins.smartpdf.SmartPdf;
2024-03-31 01:32:37 +01:00
2024-04-27 12:47:49 +02:00
constructor(optionsArg: IOpenaiProviderOptions) {
2024-03-31 01:32:37 +01:00
super();
2024-04-27 12:47:49 +02:00
this.options = optionsArg;
2024-03-31 01:32:37 +01:00
}
2024-04-27 12:47:49 +02:00
public async start() {
2024-03-31 01:32:37 +01:00
this.openAiApiClient = new plugins.openai.default({
2024-04-27 12:47:49 +02:00
apiKey: this.options.openaiToken,
2024-03-31 01:32:37 +01:00
dangerouslyAllowBrowser: true,
});
2024-04-27 12:47:49 +02:00
this.smartpdfInstance = new plugins.smartpdf.SmartPdf();
2024-03-31 01:32:37 +01:00
}
2024-04-27 12:47:49 +02:00
public async stop() {}
2024-03-31 01:32:37 +01:00
public async chatStream(input: ReadableStream<Uint8Array>): Promise<ReadableStream<string>> {
// Create a TextDecoder to handle incoming chunks
const decoder = new TextDecoder();
let buffer = '';
let currentMessage: { role: string; content: string; } | null = null;
2024-03-31 01:32:37 +01:00
// Create a TransformStream to process the input
const transform = new TransformStream<Uint8Array, string>({
async transform(chunk, controller) {
buffer += decoder.decode(chunk, { stream: true });
// Try to parse complete JSON messages from the buffer
while (true) {
const newlineIndex = buffer.indexOf('\n');
if (newlineIndex === -1) break;
const line = buffer.slice(0, newlineIndex);
buffer = buffer.slice(newlineIndex + 1);
if (line.trim()) {
try {
const message = JSON.parse(line);
currentMessage = {
role: message.role || 'user',
content: message.content || '',
};
} catch (e) {
console.error('Failed to parse message:', e);
}
}
}
// If we have a complete message, send it to OpenAI
if (currentMessage) {
const stream = await this.openAiApiClient.chat.completions.create({
model: 'gpt-4',
messages: [{ role: currentMessage.role, content: currentMessage.content }],
stream: true,
});
// Process each chunk from OpenAI
for await (const chunk of stream) {
const content = chunk.choices[0]?.delta?.content;
if (content) {
controller.enqueue(content);
}
}
currentMessage = null;
}
},
flush(controller) {
if (buffer) {
try {
const message = JSON.parse(buffer);
controller.enqueue(message.content || '');
} catch (e) {
console.error('Failed to parse remaining buffer:', e);
}
}
}
});
// Connect the input to our transform stream
return input.pipeThrough(transform);
2024-03-31 01:32:37 +01:00
}
// Implementing the synchronous chat interaction
2024-04-27 12:47:49 +02:00
public async chat(optionsArg: {
systemMessage: string;
userMessage: string;
messageHistory: {
role: 'assistant' | 'user';
content: string;
}[];
}) {
2024-03-31 01:32:37 +01:00
const result = await this.openAiApiClient.chat.completions.create({
2024-05-17 16:25:22 +02:00
model: 'gpt-4o',
2024-04-25 10:49:07 +02:00
2024-03-31 01:32:37 +01:00
messages: [
2024-04-25 10:49:07 +02:00
{ role: 'system', content: optionsArg.systemMessage },
...optionsArg.messageHistory,
{ role: 'user', content: optionsArg.userMessage },
2024-03-31 01:32:37 +01:00
],
});
return {
role: result.choices[0].message.role as 'assistant',
message: result.choices[0].message.content,
2024-03-31 01:32:37 +01:00
};
}
2024-04-25 10:49:07 +02:00
public async audio(optionsArg: { message: string }): Promise<NodeJS.ReadableStream> {
const done = plugins.smartpromise.defer<NodeJS.ReadableStream>();
2024-03-31 01:32:37 +01:00
const result = await this.openAiApiClient.audio.speech.create({
model: 'tts-1-hd',
2024-04-25 10:49:07 +02:00
input: optionsArg.message,
2024-03-31 01:32:37 +01:00
voice: 'nova',
response_format: 'mp3',
speed: 1,
});
2024-04-25 10:49:07 +02:00
const stream = result.body;
done.resolve(stream);
2024-03-31 01:32:37 +01:00
return done.promise;
}
2024-04-25 10:49:07 +02:00
public async document(optionsArg: {
2024-04-27 12:47:49 +02:00
systemMessage: string;
userMessage: string;
pdfDocuments: Uint8Array[];
2024-04-25 10:49:07 +02:00
messageHistory: {
role: 'assistant' | 'user';
content: any;
}[];
}) {
2024-04-27 12:47:49 +02:00
let pdfDocumentImageBytesArray: Uint8Array[] = [];
for (const pdfDocument of optionsArg.pdfDocuments) {
const documentImageArray = await this.smartpdfInstance.convertPDFToPngBytes(pdfDocument);
pdfDocumentImageBytesArray = pdfDocumentImageBytesArray.concat(documentImageArray);
}
console.log(`image smartfile array`);
console.log(pdfDocumentImageBytesArray.map((smartfile) => smartfile.length));
const smartfileArray = await plugins.smartarray.map(
pdfDocumentImageBytesArray,
async (pdfDocumentImageBytes) => {
return plugins.smartfile.SmartFile.fromBuffer(
'pdfDocumentImage.jpg',
Buffer.from(pdfDocumentImageBytes)
);
}
);
2024-04-25 10:49:07 +02:00
const result = await this.openAiApiClient.chat.completions.create({
2024-05-17 16:25:22 +02:00
model: 'gpt-4o',
2024-04-27 12:47:49 +02:00
// response_format: { type: "json_object" }, // not supported for now
2024-04-25 10:49:07 +02:00
messages: [
{ role: 'system', content: optionsArg.systemMessage },
...optionsArg.messageHistory,
2024-04-27 12:47:49 +02:00
{
role: 'user',
content: [
{ type: 'text', text: optionsArg.userMessage },
...(() => {
const returnArray = [];
for (const imageBytes of pdfDocumentImageBytesArray) {
returnArray.push({
type: 'image_url',
image_url: {
url: 'data:image/png;base64,' + Buffer.from(imageBytes).toString('base64'),
},
});
}
return returnArray;
})(),
],
},
2024-04-25 10:49:07 +02:00
],
});
return {
message: result.choices[0].message,
};
}
2024-03-31 01:32:37 +01:00
}