Added new json output support

src/open-ai.ts

@@ -3,15 +3,16 @@ import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean} from '@zti
 import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
 import {LLMProvider} from './provider.ts';
+import {convertSchema} from './tools.ts';
 
 export class OpenAi extends LLMProvider {
 	client!: openAI;
 
-	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string, public model: string, public llama?: boolean) {
 		super();
 		this.client = new openAI(clean({
 			baseURL: host,
-			apiKey: token || host ? 'ignored' : undefined
+			apiKey: token || (host ? 'ignored' : undefined)
 		}));
 	}
 
@@ -64,7 +65,7 @@ export class OpenAi extends LLMProvider {
 		}, [] as any[]);
 	}
 
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
 			if(options.system) {
@@ -77,8 +78,8 @@ export class OpenAi extends LLMProvider {
 				model: options.model || this.model,
 				messages: history,
 				stream: !!options.stream,
-				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
-				temperature: options.temperature || this.ai.options.llm?.temperature || 0.7,
+				max_completion_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || undefined,
+				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
 				tools: tools.map(t => ({
 					type: 'function',
 					function: {
@@ -93,6 +94,26 @@ export class OpenAi extends LLMProvider {
 				}))
 			};
 
+			if(options.schema) {
+				const schema = convertSchema(options.schema);
+				if(this.llama) {
+					delete requestParams.tools;
+					requestParams.response_format = {
+						type: 'json_schema',
+						json_schema: {name: 'json', schema}
+					}
+				} else {
+					requestParams.response_format = {
+						type: 'json_schema',
+						json_schema: {
+							name: 'response',
+							strict: true,
+							schema
+						}
+					};
+				}
+			}
+
 			let resp: any, isFirstMessage = true;
 			do {
 				resp = await this.client.chat.completions.create(requestParams).catch(err => {
@@ -158,12 +179,17 @@ export class OpenAi extends LLMProvider {
 					requestParams.messages = history;
 				}
 			} while (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
-			history.push({role: 'assistant', content: resp.choices[0].message.content.trim() || ''});
+
+			const textContent = resp.choices[0].message.content?.trim() || '';
+			history.push({role: 'assistant', content: textContent});
 			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
 			if(options.history) options.history.splice(0, options.history.length, ...history);
-			res(history.at(-1)?.content);
+
+			// Return parsed JSON if schema provided
+			const finalContent = history.at(-1)?.content;
+			res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
 		}), {abort: () => controller.abort()});
 	}
 }