feat(core): cite source documents in the AI answer (#9863)

Support issue [BS-2424](https://linear.app/affine-design/issue/BS-2424).

### What changed?
- Add relevant document prompt templates.
- Add citation rules in system prompts.
- Change message `params` type to `Record<string, any>`
- Add unit test.

<div class='graphite__hidden'>
          <div>🎥 Video uploaded on Graphite:</div>
            <a href="https://app.graphite.dev/media/video/sJGviKxfE3Ap685cl5bj/ec24e664-039e-4fab-bd26-b3312f011daf.mov">
              <img src="https://app.graphite.dev/api/v1/graphite/video/thumbnail/sJGviKxfE3Ap685cl5bj/ec24e664-039e-4fab-bd26-b3312f011daf.mov">
            </a>
          </div>
<video src="https://graphite-user-uploaded-assets-prod.s3.amazonaws.com/sJGviKxfE3Ap685cl5bj/ec24e664-039e-4fab-bd26-b3312f011daf.mov">录屏2025-01-23 10.40.38.mov</video>

packages/frontend/core/src/blocksuite/presets/ai/actions/types.ts

@@ -2,6 +2,8 @@ import type { getCopilotHistoriesQuery, RequestOptions } from '@affine/graphql';
 import type { EditorHost } from '@blocksuite/affine/block-std';
 import type { BlockModel } from '@blocksuite/affine/store';
 
+import type { DocContext } from '../chat-panel/chat-context';
+
 export const translateLangs = [
   'English',
   'Spanish',
@@ -37,7 +39,7 @@ export const imageProcessingTypes = [
 ] as const;
 
 declare global {
-  // eslint-disable-next-line @typescript-eslint/no-namespace
+  // oxlint-disable-next-line @typescript-eslint/no-namespace
   namespace BlockSuitePresets {
     type TrackerControl =
       | 'format-bar'
@@ -57,6 +59,7 @@ declare global {
     }
 
     interface AITextActionOptions {
+      // user input text
       input?: string;
       stream?: boolean;
       attachments?: (string | File | Blob)[]; // blob could only be strings for the moments (url or data urls)
@@ -101,6 +104,8 @@ declare global {
       T['stream'] extends true ? TextStream : Promise<string>;
 
     interface ChatOptions extends AITextActionOptions {
+      // related documents
+      docs?: DocContext[];
       sessionId?: string;
       isRootSession?: boolean;
     }

packages/frontend/core/src/blocksuite/presets/ai/chat-panel/chat-context.ts

@@ -76,7 +76,7 @@ export interface BaseChip {
 
 export interface DocChip extends BaseChip {
   docId: string;
-  content?: Signal<string>;
+  markdown?: Signal<string>;
 }
 
 export interface FileChip extends BaseChip {

packages/frontend/core/src/blocksuite/presets/ai/chat-panel/chat-panel-input.ts

@@ -25,7 +25,7 @@ import { AIProvider } from '../provider';
 import { reportResponse } from '../utils/action-reporter';
 import { readBlobAsURL } from '../utils/image';
 import type { AINetworkSearchConfig } from './chat-config';
-import type { ChatContextValue, ChatMessage } from './chat-context';
+import type { ChatContextValue, ChatMessage, DocContext } from './chat-context';
 import { isDocChip } from './components/utils';
 
 const MaximumImageCount = 32;
@@ -512,16 +512,11 @@ export class ChatPanelInput extends SignalWatcher(WithDisposable(LitElement)) {
     if (status === 'loading' || status === 'transmitting') return;
 
     const { images } = this.chatContextValue;
-    if (!text && images.length === 0) {
+    if (!text) {
       return;
     }
     const { doc } = this.host;
 
-    const docsContent = chips
-      .filter(isDocChip)
-      .map(chip => chip.content?.value || '')
-      .join('\n');
-
     this.updateContext({
       images: [],
       status: 'loading',
@@ -534,16 +529,14 @@ export class ChatPanelInput extends SignalWatcher(WithDisposable(LitElement)) {
       images?.map(image => readBlobAsURL(image))
     );
 
-    const content =
-      (markdown ? `${markdown}\n` : '') + `${docsContent}\n` + text;
-
+    const userInput = (markdown ? `${markdown}\n` : '') + text;
     this.updateContext({
       items: [
         ...this.chatContextValue.items,
         {
           id: '',
           role: 'user',
-          content: content,
+          content: userInput,
           createdAt: new Date().toISOString(),
           attachments,
         },
@@ -558,8 +551,16 @@ export class ChatPanelInput extends SignalWatcher(WithDisposable(LitElement)) {
 
     try {
       const abortController = new AbortController();
+      const docs: DocContext[] = chips
+        .filter(isDocChip)
+        .filter(chip => !!chip.markdown?.value && chip.state === 'success')
+        .map(chip => ({
+          docId: chip.docId,
+          markdown: chip.markdown?.value || '',
+        }));
       const stream = AIProvider.actions.chat?.({
-        input: content,
+        input: userInput,
+        docs: docs,
         docId: doc.id,
         attachments: images,
         workspaceId: doc.workspace.id,

packages/frontend/core/src/blocksuite/presets/ai/chat-panel/components/doc-chip.ts

@@ -100,10 +100,10 @@ export class ChatPanelDocChip extends SignalWatcher(
         doc.load();
       }
       const result = await extractMarkdownFromDoc(doc, this.host.std.provider);
-      if (this.chip.content) {
-        this.chip.content.value = result.markdown;
+      if (this.chip.markdown) {
+        this.chip.markdown.value = result.markdown;
       } else {
-        this.chip.content = new Signal<string>(result.markdown);
+        this.chip.markdown = new Signal<string>(result.markdown);
       }
       this.updateChip(this.chip, {
         state: 'success',

packages/frontend/core/src/components/blocksuite/block-suite-editor/ai/request.ts

@@ -17,7 +17,7 @@ export type TextToTextOptions = {
   sessionId?: string | Promise<string>;
   content?: string;
   attachments?: (string | Blob | File)[];
-  params?: Record<string, string>;
+  params?: Record<string, any>;
   timeout?: number;
   stream?: boolean;
   signal?: AbortSignal;

packages/frontend/core/src/components/blocksuite/block-suite-editor/ai/setup-provider.tsx

@@ -117,11 +117,22 @@ export function setupAIProvider(
     const sessionId =
       options.sessionId ??
       getChatSessionId(options.workspaceId, options.docId, options.attachments);
+    const { input, docs, ...rest } = options;
+    const params = docs?.length
+      ? {
+          docs: docs.map((doc, i) => ({
+            docId: doc.docId,
+            markdown: doc.markdown,
+            index: i + 1,
+          })),
+        }
+      : undefined;
     return textToText({
-      ...options,
+      ...rest,
       client,
-      content: options.input,
+      content: input,
       sessionId,
+      params,
     });
   });