263 lines
11 KiB
TypeScript
263 lines
11 KiB
TypeScript
import express from "express";
|
||
import u from "@/utils";
|
||
import { z } from "zod";
|
||
import { error, success } from "@/lib/responseFormat";
|
||
import { validateFields } from "@/middleware/middleware";
|
||
import { useSkill } from "@/utils/agent/skillsTools";
|
||
import { tool } from "ai";
|
||
import { o_script } from "@/types/database";
|
||
|
||
const router = express.Router();
|
||
|
||
/** 新资产：AI 首次识别到的资产，需要完整信息 */
|
||
const NewAssetSchema = z.object({
|
||
name: z.string().describe("资产名称,仅为名称不做其他任何表述"),
|
||
desc: z.string().describe("资产描述"),
|
||
type: z.enum(["role", "tool", "scene"]).describe("资产类型"),
|
||
scriptIds: z.array(z.number()).describe("使用该资产的剧本id数组"),
|
||
});
|
||
|
||
/** 已有资产：数据库中已存在的资产，只需给出名称和关联的剧本 */
|
||
const ExistingAssetRefSchema = z.object({
|
||
name: z.string().describe("已有资产的名称,必须与已有资产列表中的名称完全一致"),
|
||
scriptIds: z.array(z.number()).describe("使用该资产的剧本id数组"),
|
||
});
|
||
|
||
export const AssetSchema = z.object({
|
||
name: z.string().describe("资产名称,仅为名称不做其他任何表述"),
|
||
desc: z.string().describe("资产描述"),
|
||
type: z.enum(["role", "tool", "scene"]).describe("资产类型"),
|
||
});
|
||
|
||
type NewAsset = z.infer<typeof NewAssetSchema>;
|
||
type ExistingAssetRef = z.infer<typeof ExistingAssetRefSchema>;
|
||
type Asset = z.infer<typeof AssetSchema>;
|
||
|
||
/** 每批 AI 调用的结果 */
|
||
type GroupResult = {
|
||
batchScriptIds: number[];
|
||
newAssets: NewAsset[];
|
||
existingRefs: ExistingAssetRef[];
|
||
} | null;
|
||
|
||
/** 将 scriptIds 数组按 groupSize 分组 */
|
||
function chunkArray<T>(arr: T[], groupSize: number): T[][] {
|
||
const chunks: T[][] = [];
|
||
for (let i = 0; i < arr.length; i += groupSize) {
|
||
chunks.push(arr.slice(i, i + groupSize));
|
||
}
|
||
return chunks;
|
||
}
|
||
|
||
export default router.post(
|
||
"/",
|
||
validateFields({
|
||
scriptIds: z.array(z.number()),
|
||
projectId: z.number(),
|
||
groupSize: z.number().min(1).max(10).optional(),
|
||
}),
|
||
async (req, res) => {
|
||
const { scriptIds, projectId, groupSize = 5 } = req.body;
|
||
if (!scriptIds.length) return res.status(400).send(error("请先选择剧本"));
|
||
const scripts = await u.db("o_script").whereIn("id", scriptIds);
|
||
const intansce = u.Ai.Text("universalAi");
|
||
|
||
// 查询已有的剧本-资产关联，找出已经提取过资产的剧本
|
||
const existingScriptAssets = await u.db("o_scriptAssets").whereIn("scriptId", scriptIds).select("scriptId");
|
||
const scriptIdsWithAssets = new Set(existingScriptAssets.map((sa: any) => sa.scriptId));
|
||
|
||
// 构建 scriptId -> script 内容的映射
|
||
const scriptMap = new Map(scripts.map((s: o_script) => [s.id, s]));
|
||
|
||
// 过滤掉已成功提取过资产的剧本（extractState === 1 且有关联资产）
|
||
const filteredScriptIds = scriptIds.filter((id: number) => {
|
||
const script = scriptMap.get(id);
|
||
return !(script?.extractState === 1 && scriptIdsWithAssets.has(id));
|
||
});
|
||
const skippedCount = scriptIds.length - filteredScriptIds.length;
|
||
|
||
if (!filteredScriptIds.length) {
|
||
return res.send(success("所有剧本已提取过资产，无需重复提取"));
|
||
}
|
||
|
||
await u.db("o_script").whereIn("id", filteredScriptIds).update({
|
||
extractState: 0,
|
||
});
|
||
|
||
const errors: { scriptId: number; error: string }[] = [];
|
||
let successCount = 0;
|
||
|
||
// 将过滤后的 scriptIds 按 groupSize（默认5）分组，每组一起发给 AI
|
||
const scriptGroups = chunkArray(filteredScriptIds, groupSize);
|
||
|
||
/** 一组剧本提取完成后统一入库并建立关联 */
|
||
async function persistGroupResult(result: GroupResult) {
|
||
if (!result) return;
|
||
const { batchScriptIds, newAssets, existingRefs } = result;
|
||
if (!newAssets.length && !existingRefs.length) return;
|
||
|
||
// 查询已有资产
|
||
const existingAssets = await u.db("o_assets").where("projectId", projectId).select("id", "name");
|
||
const existingMap = new Map(existingAssets.map((a) => [a.name!, a.id!]));
|
||
|
||
// 插入新资产（不在已有列表中的）
|
||
const toInsert = newAssets.filter((asset) => !existingMap.has(asset.name));
|
||
if (toInsert.length) {
|
||
await u.db("o_assets").insert(
|
||
toInsert.map((asset) => ({
|
||
name: asset.name,
|
||
type: asset.type,
|
||
describe: asset.desc,
|
||
projectId: projectId,
|
||
startTime: Date.now(),
|
||
})),
|
||
);
|
||
}
|
||
|
||
// 重新查询获取完整的 name -> id 映射
|
||
const allAssets = await u.db("o_assets").where("projectId", projectId).select("id", "name");
|
||
const nameToId = new Map(allAssets.map((a) => [a.name, a.id]));
|
||
|
||
// 收集所有资产与剧本的关联关系
|
||
const scriptAssetRows: { scriptId: number; assetId: number }[] = [];
|
||
|
||
// 新资产的关联
|
||
for (const asset of newAssets) {
|
||
const assetId = nameToId.get(asset.name);
|
||
if (assetId) {
|
||
for (const sid of asset.scriptIds) {
|
||
scriptAssetRows.push({ scriptId: sid, assetId });
|
||
}
|
||
}
|
||
}
|
||
|
||
// 已有资产的关联
|
||
for (const ref of existingRefs) {
|
||
const assetId = nameToId.get(ref.name);
|
||
if (assetId) {
|
||
for (const sid of ref.scriptIds) {
|
||
scriptAssetRows.push({ scriptId: sid, assetId });
|
||
}
|
||
}
|
||
}
|
||
|
||
// 先删除本批 scriptId 的旧关联，再插入新的
|
||
await u.db("o_scriptAssets").whereIn("scriptId", batchScriptIds).delete();
|
||
if (scriptAssetRows.length) {
|
||
await u.db("o_scriptAssets").insert(scriptAssetRows);
|
||
}
|
||
|
||
// 本批成功的剧本状态更新为 1（成功）
|
||
await u.db("o_script").whereIn("id", batchScriptIds).update({
|
||
extractState: 1,
|
||
errorReason: null,
|
||
});
|
||
}
|
||
|
||
// 逐组处理（每组最多 groupSize 集剧本一起发给 AI）
|
||
for (const group of scriptGroups) {
|
||
// 过滤有效剧本
|
||
const validScripts: { id: number; script: o_script }[] = [];
|
||
for (const scriptId of group as number[]) {
|
||
const script = scriptMap.get(scriptId);
|
||
if (!script) {
|
||
errors.push({ scriptId, error: "未找到对应剧本" });
|
||
await u.db("o_script").where("id", scriptId).update({ extractState: -1, errorReason: "未找到对应剧本" });
|
||
} else {
|
||
validScripts.push({ id: scriptId, script });
|
||
}
|
||
}
|
||
if (!validScripts.length) continue;
|
||
|
||
// 查询当前项目已有的资产列表，提供给 AI 参考
|
||
const existingAssets = await u.db("o_assets").where("projectId", projectId).select("name", "type");
|
||
const existingAssetsList = existingAssets.map((a) => `${a.name}(${a.type})`).join("、");
|
||
|
||
// 拼接多集剧本内容，每集用分隔标记
|
||
const scriptsContent = validScripts
|
||
.map(({ id, script }) => `===== 【剧本ID: ${id}】${script.name || ""} =====\n${script.content}`)
|
||
.join("\n\n");
|
||
|
||
const validScriptIds = validScripts.map((v) => v.id);
|
||
|
||
// 用闭包收集 AI 返回的资产
|
||
let collectedNew: NewAsset[] = [];
|
||
let collectedExisting: ExistingAssetRef[] = [];
|
||
|
||
const resultTool = tool({
|
||
description: "返回结果时必须调用这个工具",
|
||
inputSchema: z.object({
|
||
newAssets: z
|
||
.array(NewAssetSchema)
|
||
.describe("新发现的资产列表（不在已有资产列表中的），需要完整的 prompt、name、desc、type 和使用该资产的 scriptIds"),
|
||
existingAssetRefs: z
|
||
.array(ExistingAssetRefSchema)
|
||
.describe("已有资产的引用列表（在已有资产列表中已存在的），只需给出资产名称和使用该资产的 scriptIds"),
|
||
}),
|
||
execute: async ({ newAssets, existingAssetRefs }) => {
|
||
console.log("[tools] extractAssets result", { newAssets, existingAssetRefs });
|
||
if (newAssets?.length) collectedNew = newAssets;
|
||
if (existingAssetRefs?.length) collectedExisting = existingAssetRefs;
|
||
return "无需回复用户任何内容";
|
||
},
|
||
});
|
||
|
||
try {
|
||
const data = await u.db("o_prompt").where("type", "scriptAssetExtraction").first("data");
|
||
const existingHint = existingAssetsList
|
||
? `\n\n【已有资产列表】：${existingAssetsList}\n对于已有资产，如果在剧本中出现，只需在 existingAssetRefs 中给出资产名称和对应的 scriptIds 数组即可，无需重复生成 prompt/desc/type。对于新发现的资产（不在已有列表中），请在 newAssets 中给出完整信息。`
|
||
: "";
|
||
|
||
const output = await intansce.invoke({
|
||
messages: [
|
||
{
|
||
role: "system",
|
||
content:
|
||
data?.data +
|
||
"\n\n提取剧本中涉及的资产（角色、场景、道具），参考技能 script_assets_extract 规范，结果必须通过 resultTool 工具返回。" +
|
||
"\n\n注意：本次会同时提供多集剧本，每集剧本以 ===== 【剧本ID: xxx】 ===== 分隔。你需要分析每集剧本使用了哪些资产，并在输出中用 scriptIds 数组标明每个资产在哪些剧本中出现。",
|
||
},
|
||
{
|
||
role: "user",
|
||
content: `当前已有资产列表：${existingHint}\n\n请根据以下${validScripts.length}集剧本提取对应的剧本资产（角色、场景、道具）:\n\n${scriptsContent}`,
|
||
},
|
||
],
|
||
tools: { resultTool },
|
||
});
|
||
console.log("%c Line:extractAssets 🍧 output", "background:#f5ce50", output.text);
|
||
} catch (e: any) {
|
||
const msg = e?.message || String(e);
|
||
const scriptNames = validScripts.map((v) => v.script.name).join(", ");
|
||
console.error(`[extractAssets] group=[${validScriptIds.join(",")}] 提取失败:`, msg);
|
||
for (const { id, script } of validScripts) {
|
||
errors.push({ scriptId: id, error: (script.name || "") + ":" + u.error(e).message });
|
||
await u
|
||
.db("o_script")
|
||
.where("id", id)
|
||
.update({ extractState: -1, errorReason: u.error(e).message });
|
||
}
|
||
continue;
|
||
}
|
||
|
||
if (!collectedNew.length && !collectedExisting.length) {
|
||
for (const { id } of validScripts) {
|
||
errors.push({ scriptId: id, error: "AI 未返回任何资产" });
|
||
await u.db("o_script").where("id", id).update({ extractState: -1, errorReason: "AI 未返回任何资产" });
|
||
}
|
||
continue;
|
||
}
|
||
|
||
successCount += validScripts.length;
|
||
|
||
// 入库
|
||
await persistGroupResult({
|
||
batchScriptIds: validScriptIds,
|
||
newAssets: collectedNew,
|
||
existingRefs: collectedExisting,
|
||
});
|
||
}
|
||
|
||
return res.send(success(skippedCount > 0 ? `开始提取资产，跳过 ${skippedCount} 个已提取的剧本` : "开始提取资产"));
|
||
},
|
||
);
|