feat(core): add o4-mini model (#12175)

Close [AI-85](https://linear.app/affine-design/issue/AI-85)

<!-- This is an auto-generated comment: release notes by coderabbit.ai -->
## Summary by CodeRabbit

- **New Features**
	- Enhanced AI chat experience with improved web search integration, including a new "AUTO" mode that allows the AI to search the web only when needed.
	- Updated AI model for chat to "o4-mini" for improved performance.
- **Improvements**
	- Refined instructions to reduce hallucinations and ensure the AI admits uncertainty when unsure.
	- Enhanced streaming responses to display web search queries and results more clearly.
	- Updated web search tool modes for more accurate and flexible search behavior.
	- Centralized AI provider options and improved tool selection logic for better response quality.
- **Bug Fixes**
	- Improved handling of tool calls and reasoning steps in AI chat responses.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
This commit is contained in:
akumatus
2025-05-07 13:36:41 +00:00
parent b4f56c0a78
commit 32db367298
7 changed files with 105 additions and 32 deletions

View File

@@ -1037,24 +1037,26 @@ Finally, please only send us the content of your continuation in Markdown Format
const chat: Prompt[] = [
{
name: 'Chat With AFFiNE AI',
model: 'gpt-4.1',
model: 'o4-mini',
messages: [
{
role: 'system',
content: `You are AFFiNE AI, a professional and humorous copilot within AFFiNE. You are powered by latest GPT model from OpenAI and AFFiNE. AFFiNE is an open source general purposed productivity tool that contains unified building blocks that users can use on any interfaces, including block-based docs editor, infinite canvas based edgeless graphic mode, or multi-dimensional table with multiple transformable views. Your mission is always to try your very best to assist users to use AFFiNE to write docs, draw diagrams or plan things with these abilities. You always think step-by-step and describe your plan for what to build, using well-structured and clear markdown, written out in great detail. Unless otherwise specified, where list, JSON, or code blocks are required for giving the output. Minimize any other prose so that your responses can be directly used and inserted into the docs. You are able to access to API of AFFiNE to finish your job. You always respect the users' privacy and would not leak their info to anyone else. AFFiNE is made by Toeverything .Pte .Ltd, a company registered in Singapore with a diverse and international team. The company also open sourced blocksuite and octobase for building tools similar to Affine. The name AFFiNE comes from the idea of AFFiNE transform, as blocks in affine can all transform in page, edgeless or database mode. AFFiNE team is now having 25 members, an open source company driven by engineers. Today is: {{affine::date}}, User's preferred language is {{affine::language}}.
# Response Guide
Use the web search tool to gather information from the web. There are two modes for web searching:
Use the web search tool to gather information from the web if you have been equipped with it. There are two modes for web searching:
- MUST: Means you always need to use the web search tool to gather information from the web, no matter what the user's query is.
- CAN: Indicates that web searching is optional - you may use the web search tool at your discretion when you determine it would provide valuable information for answering the user's query.
- AUTO: Indicates that web searching is optional - you may use the web search tool at your discretion when you determine it would provide valuable information for answering the user's query. If your own knowledge can directly answer the user's questions, there is no need to use web search tool.
Currently, you are in the {{searchMode}} web searching mode.
I will provide you with some content fragments. There are two types of content fragments:
- Document fragments, identified by a \`document_id\` and containing \`document_content\`.
- File fragments, identified by a \`blob_id\` and containing \`file_content\`.
You need to analyze web search results and content fragments, determine their relevance to the user's query, and combine them to answer the user's query.
Please cite all source links in your final answer according to the citations rules.
You need to analyze web search results and content fragments, determine their relevance to the user's query, and combine them with your own knowledge to answer the user's query.
Please cite all source links in your final answer according to the citations rules. Don't make up citations that don't exist.
If youre unsure, tell them youre unsure so they dont fall into hallucinations.
## Citations Rules
When referencing information from the provided documents, files or web search results in your response:

View File

@@ -177,6 +177,12 @@ export class AnthropicProvider
yield message.textDelta;
break;
}
case 'tool-call': {
if (message.toolName === 'web_search') {
yield '\n' + `Searching the web "${message.args.query}"` + '\n';
}
break;
}
case 'tool-result': {
if (message.toolName === 'web_search') {
this.toolResults.push(this.getWebSearchLinks(message.result));
@@ -219,18 +225,15 @@ export class AnthropicProvider
return undefined;
}
private getAnthropicOptions(
options: CopilotChatOptions
): AnthropicProviderOptions {
private getAnthropicOptions(options: CopilotChatOptions) {
const result: AnthropicProviderOptions = {};
if (options?.reasoning) {
return {
thinking: {
type: 'enabled',
budgetTokens: 12000,
},
result.thinking = {
type: 'enabled',
budgetTokens: 12000,
};
}
return {};
return result;
}
private getWebSearchLinks(

View File

@@ -2,6 +2,7 @@ import {
createOpenAI,
openai,
type OpenAIProvider as VercelOpenAIProvider,
OpenAIResponsesProviderOptions,
} from '@ai-sdk/openai';
import {
AISDKError,
@@ -10,7 +11,6 @@ import {
generateObject,
generateText,
streamText,
ToolSet,
} from 'ai';
import {
@@ -19,6 +19,7 @@ import {
metrics,
UserFriendlyError,
} from '../../../base';
import { createExaTool } from '../tools';
import { CopilotProvider } from './provider';
import {
ChatMessageRole,
@@ -42,6 +43,11 @@ export type OpenAIConfig = {
baseUrl?: string;
};
type OpenAITools = {
web_search_preview: ReturnType<typeof openai.tools.webSearchPreview>;
web_search: ReturnType<typeof createExaTool>;
};
export class OpenAIProvider
extends CopilotProvider<OpenAIConfig>
implements
@@ -68,7 +74,7 @@ export class OpenAIProvider
'gpt-4.1-2025-04-14',
'gpt-4.1-mini',
'o1',
'o3-mini',
'o4-mini',
// embeddings
'text-embedding-3-large',
'text-embedding-3-small',
@@ -81,6 +87,8 @@ export class OpenAIProvider
'gpt-image-1',
];
private readonly MAX_STEPS = 20;
#instance!: VercelOpenAIProvider;
override configured(): boolean {
@@ -179,21 +187,28 @@ export class OpenAIProvider
}
}
private getTools(options: CopilotChatOptions): ToolSet | undefined {
private getTools(
options: CopilotChatOptions,
model: string
): Partial<OpenAITools> {
const tools: Partial<OpenAITools> = {};
if (options?.tools?.length) {
const tools: ToolSet = {};
for (const tool of options.tools) {
switch (tool) {
case 'webSearch': {
tools.web_search_preview = openai.tools.webSearchPreview();
// o series reasoning models
if (model.startsWith('o')) {
tools.web_search = createExaTool(this.AFFiNEConfig);
} else {
tools.web_search_preview = openai.tools.webSearchPreview();
}
break;
}
}
}
return tools;
}
return undefined;
return tools;
}
// ====== text to text ======
@@ -231,10 +246,10 @@ export class OpenAIProvider
: await generateText({
...commonParams,
providerOptions: {
openai: options.user ? { user: options.user } : {},
openai: this.getOpenAIOptions(options),
},
toolChoice: options.webSearch ? 'required' : 'auto',
tools: this.getTools(options),
tools: this.getTools(options, model),
maxSteps: this.MAX_STEPS,
});
return text.trim();
@@ -258,14 +273,16 @@ export class OpenAIProvider
const modelInstance = this.#instance.responses(model);
const tools = this.getTools(options, model);
const { fullStream } = streamText({
model: modelInstance,
system,
messages: msgs,
tools: this.getTools(options),
providerOptions: {
openai: options.user ? { user: options.user } : {},
openai: this.getOpenAIOptions(options),
},
tools: tools as OpenAITools,
maxSteps: this.MAX_STEPS,
frequencyPenalty: options.frequencyPenalty || 0,
presencePenalty: options.presencePenalty || 0,
temperature: options.temperature || 0,
@@ -274,25 +291,53 @@ export class OpenAIProvider
});
const parser = new CitationParser();
let lastType;
for await (const chunk of fullStream) {
if (chunk) {
switch (chunk.type) {
case 'text-delta': {
if (lastType !== chunk.type) {
yield `\n`;
}
const result = parser.parse(chunk.textDelta);
yield result;
break;
}
case 'reasoning': {
if (lastType !== chunk.type) {
yield `\n`;
}
yield chunk.textDelta;
break;
}
case 'tool-call': {
if (chunk.toolName === 'web_search') {
yield '\n' + `Searching the web "${chunk.args.query}"` + '\n';
}
break;
}
case 'tool-result': {
if (chunk.toolName === 'web_search') {
yield '\n' + this.getWebSearchLinks(chunk.result) + '\n';
}
break;
}
case 'step-finish': {
const result = parser.end();
yield result;
break;
}
case 'error': {
const error = chunk.error as { type: string; message: string };
throw new Error(error.message);
}
}
if (options.signal?.aborted) {
await fullStream.cancel();
break;
}
lastType = chunk.type;
}
}
} catch (e: any) {
@@ -380,4 +425,28 @@ export class OpenAIProvider
throw e;
}
}
private getOpenAIOptions(options: CopilotChatOptions) {
const result: OpenAIResponsesProviderOptions = {};
if (options?.reasoning) {
result.reasoningEffort = 'medium';
result.reasoningSummary = 'auto';
}
if (options?.user) {
result.user = options.user;
}
return result;
}
private getWebSearchLinks(
list: {
title: string | null;
url: string;
}[]
): string {
const links = list.reduce((acc, result) => {
return acc + `\n[${result.title ?? result.url}](${result.url})\n`;
}, '\n');
return links + '\n';
}
}

View File

@@ -10,8 +10,7 @@ export const createExaTool = (config: Config) => {
parameters: z.object({
query: z.string().describe('The query to search the web for.'),
mode: z
.enum(['MUST', 'CAN'])
.optional()
.enum(['MUST', 'AUTO'])
.describe('The mode to search the web for.'),
}),
execute: async ({ query, mode }) => {