162 lines
5.8 KiB
TypeScript
162 lines
5.8 KiB
TypeScript
import * as z from "zod";
|
||
import { ModelMessage, Output } from "ai";
|
||
|
||
import { o_novel } from "@/types/database";
|
||
import ai from "@/utils/ai";
|
||
import u from "@/utils";
|
||
export interface EventType {
|
||
name: string;
|
||
detail: string;
|
||
chapter: string;
|
||
}
|
||
|
||
export interface Novel {
|
||
event: EventType[];
|
||
}
|
||
// 章节拆分
|
||
function getChapterGroups<T>(chapters: T[], windowSize: number = 5, overlap: number = 1): T[][] {
|
||
const res: T[][] = [];
|
||
if (windowSize < 1 || overlap < 0) return res;
|
||
let i = 0;
|
||
const length = chapters.length;
|
||
while (i < length) {
|
||
if (res.length === 0) {
|
||
// 第一组，直接取 windowSize 个
|
||
res.push(chapters.slice(i, i + windowSize));
|
||
i += windowSize;
|
||
} else {
|
||
// 取上一组最后 overlap 个，加上新的 windowSize 个
|
||
const prevGroup = res[res.length - 1];
|
||
const overlapItems = prevGroup.slice(-overlap);
|
||
const newItems = chapters.slice(i, i + windowSize);
|
||
if (newItems.length === 0) break; // 已经取完，跳出
|
||
res.push([...overlapItems, ...newItems]);
|
||
i += windowSize;
|
||
}
|
||
}
|
||
return res;
|
||
}
|
||
|
||
/* 文本数据清洗
|
||
* @param textData 需要清洗的文本
|
||
* @param windowSize 每组数量 默认5
|
||
* @param overlap 交叠数量 默认1
|
||
* @returns {totalCharacter:所有人物角色卡,totalEvent:所有事件}
|
||
*/
|
||
|
||
class CleanNovel {
|
||
windowSize: number;
|
||
overlap: number;
|
||
constructor(windowSize: number = 5, overlap: number = 1) {
|
||
this.windowSize = windowSize;
|
||
this.overlap = overlap;
|
||
}
|
||
async start(allChapters: o_novel[], projectId: number): Promise<EventType[]> {
|
||
const groups = getChapterGroups(allChapters!, this.windowSize, this.overlap);
|
||
|
||
let preData: Novel | null = null;
|
||
//所有事件
|
||
let totalEvent: EventType[] = [];
|
||
const intansce = u.Ai.Text("eventExtractAi");
|
||
|
||
try {
|
||
for (let gi = 0; gi < groups.length; gi++) {
|
||
const group = groups[gi];
|
||
// 第一批没有交叠章节，后续批次前 overlap 个是交叠章节（仅作上下文，不输出事件）
|
||
const overlapCount = gi === 0 ? 0 : this.overlap;
|
||
const overlapChapterIndexes = group.slice(0, overlapCount).map((i) => i.chapterIndex);
|
||
|
||
const cleanText = group
|
||
.map((i, index: number) => {
|
||
const isOverlap = overlapChapterIndexes.includes(i.chapterIndex);
|
||
return {
|
||
role: "user",
|
||
content: isOverlap
|
||
? `【上文衔接章节，仅供上下文参考，禁止为本章生成情节单元】\n第${i.chapterIndex}章：\n\n${i.chapterData}`
|
||
: `第${i.chapterIndex}章：\n\n${i.chapterData}`,
|
||
} as ModelMessage;
|
||
})
|
||
.filter(Boolean);
|
||
const taskRecord = await u.task(projectId, "事件提取", "gpt-4.1", {
|
||
describe: "根据小说原文，提取情节单元",
|
||
content: cleanText,
|
||
});
|
||
let resData;
|
||
try {
|
||
resData = await intansce.invoke({
|
||
messages: [
|
||
{
|
||
role: "system",
|
||
content: `
|
||
你是专业剧本结构分析师，负责将用户提供的章节文本拆分为标准情节单元。请严格遵循以下规则执行。
|
||
|
||
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
||
【情节单元拆分规则】
|
||
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
||
|
||
▍拆分粒度
|
||
- 连续多个章节若属同一戏剧动作，可合并为 1 个情节单元，禁止过细拆分；
|
||
- 每个情节单元的 detail 字数控制在 100～200 字。
|
||
|
||
▍每个情节单元包含以下三个字段
|
||
- chapter：事件覆盖的章节范围（如"1-3章"），每个章节只能归属一个事件；
|
||
- name：事件名称，须具体描述实际戏剧动作，禁止使用"XXX踏上征程""命运转折"等笼统标题；
|
||
- detail：事件过程详情，包含时间、地点、涉及人物、起因、经过、结果。
|
||
|
||
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
||
【执行规则】
|
||
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
|
||
|
||
✅ 必须执行
|
||
1. 所有章节按剧情顺序逐一覆盖，不得遗漏；
|
||
2. 标注为【上文衔接章节】的内容仅作上下文理解使用，禁止为其生成任何情节单元。
|
||
|
||
🚫 禁止出现
|
||
- 笼统事件名称；
|
||
- 单个章节拆分为多个情节单元；
|
||
- 遗漏任何章节。
|
||
`,
|
||
},
|
||
...cleanText,
|
||
],
|
||
output: Output.object({
|
||
schema: z.object({
|
||
event: z.array(
|
||
z
|
||
.object({
|
||
chapter: z
|
||
.string()
|
||
.describe(
|
||
"事件覆盖的章节（如1-3章、4-6章），章节划分必须连续，每个章节范围只能属于一个事件。事件分割不可过细——避免只描述琐碎、日常细节的微小事件。",
|
||
),
|
||
name: z.string().describe("事件名称"),
|
||
detail: z.string().describe("事件过程详情（包括起因、经过、结果、场景、人物等）"),
|
||
})
|
||
.describe("事件必须在100-200字说明起因经过结果，不可将单一章节或细小场景独立成事件，"),
|
||
),
|
||
}),
|
||
}),
|
||
});
|
||
} catch (e) {
|
||
taskRecord(-1, u.error(e).message);
|
||
throw e;
|
||
}
|
||
taskRecord(1);
|
||
|
||
preData = JSON.parse(resData.text);
|
||
|
||
const newEvents = preData?.event || [];
|
||
newEvents.forEach((newItem) => {
|
||
totalEvent.push({ ...newItem });
|
||
});
|
||
}
|
||
} catch (e) {
|
||
console.error(e);
|
||
throw e;
|
||
}
|
||
return totalEvent;
|
||
}
|
||
}
|
||
|
||
export default CleanNovel;
|