mirror of
https://github.com/EthanMarti/infio-copilot.git
synced 2026-01-16 08:21:55 +00:00
390 lines
11 KiB
TypeScript
390 lines
11 KiB
TypeScript
import { Result, err, ok } from "neverthrow";
|
||
|
||
import { LLMModel } from '../../types/llm/model';
|
||
import { RequestMessage } from '../../types/llm/request';
|
||
import { InfioSettings } from '../../types/settings';
|
||
import { tokenCount } from '../../utils/token';
|
||
import LLMManager from '../llm/manager';
|
||
import { ANALYZE_PAPER_DESCRIPTION, ANALYZE_PAPER_PROMPT } from '../prompts/transformations/analyze-paper';
|
||
import { DENSE_SUMMARY_DESCRIPTION, DENSE_SUMMARY_PROMPT } from '../prompts/transformations/dense-summary';
|
||
import { KEY_INSIGHTS_DESCRIPTION, KEY_INSIGHTS_PROMPT } from '../prompts/transformations/key-insights';
|
||
import { REFLECTIONS_DESCRIPTION, REFLECTIONS_PROMPT } from '../prompts/transformations/reflections';
|
||
import { SIMPLE_SUMMARY_DESCRIPTION, SIMPLE_SUMMARY_PROMPT } from '../prompts/transformations/simple-summary';
|
||
import { TABLE_OF_CONTENTS_DESCRIPTION, TABLE_OF_CONTENTS_PROMPT } from '../prompts/transformations/table-of-contents';
|
||
|
||
// 转换类型枚举
|
||
export enum TransformationType {
|
||
DENSE_SUMMARY = 'dense-summary',
|
||
ANALYZE_PAPER = 'analyze-paper',
|
||
SIMPLE_SUMMARY = 'simple-summary',
|
||
KEY_INSIGHTS = 'key-insights',
|
||
TABLE_OF_CONTENTS = 'table-of-contents',
|
||
REFLECTIONS = 'reflections'
|
||
}
|
||
|
||
// 转换配置接口
|
||
export interface TransformationConfig {
|
||
type: TransformationType;
|
||
prompt: string;
|
||
description: string;
|
||
maxTokens?: number;
|
||
}
|
||
|
||
// 所有可用的转换配置
|
||
export const TRANSFORMATIONS: Record<TransformationType, TransformationConfig> = {
|
||
[TransformationType.DENSE_SUMMARY]: {
|
||
type: TransformationType.DENSE_SUMMARY,
|
||
prompt: DENSE_SUMMARY_PROMPT,
|
||
description: DENSE_SUMMARY_DESCRIPTION,
|
||
maxTokens: 4000
|
||
},
|
||
[TransformationType.ANALYZE_PAPER]: {
|
||
type: TransformationType.ANALYZE_PAPER,
|
||
prompt: ANALYZE_PAPER_PROMPT,
|
||
description: ANALYZE_PAPER_DESCRIPTION,
|
||
maxTokens: 3000
|
||
},
|
||
[TransformationType.SIMPLE_SUMMARY]: {
|
||
type: TransformationType.SIMPLE_SUMMARY,
|
||
prompt: SIMPLE_SUMMARY_PROMPT,
|
||
description: SIMPLE_SUMMARY_DESCRIPTION,
|
||
maxTokens: 2000
|
||
},
|
||
[TransformationType.KEY_INSIGHTS]: {
|
||
type: TransformationType.KEY_INSIGHTS,
|
||
prompt: KEY_INSIGHTS_PROMPT,
|
||
description: KEY_INSIGHTS_DESCRIPTION,
|
||
maxTokens: 3000
|
||
},
|
||
[TransformationType.TABLE_OF_CONTENTS]: {
|
||
type: TransformationType.TABLE_OF_CONTENTS,
|
||
prompt: TABLE_OF_CONTENTS_PROMPT,
|
||
description: TABLE_OF_CONTENTS_DESCRIPTION,
|
||
maxTokens: 2000
|
||
},
|
||
[TransformationType.REFLECTIONS]: {
|
||
type: TransformationType.REFLECTIONS,
|
||
prompt: REFLECTIONS_PROMPT,
|
||
description: REFLECTIONS_DESCRIPTION,
|
||
maxTokens: 2500
|
||
}
|
||
};
|
||
|
||
// 转换参数接口
|
||
export interface TransformationParams {
|
||
content: string;
|
||
transformationType: TransformationType;
|
||
settings: InfioSettings;
|
||
model?: LLMModel;
|
||
maxContentTokens?: number;
|
||
}
|
||
|
||
// 转换结果接口
|
||
export interface TransformationResult {
|
||
success: boolean;
|
||
result?: string;
|
||
error?: string;
|
||
truncated?: boolean;
|
||
originalTokens?: number;
|
||
processedTokens?: number;
|
||
}
|
||
|
||
/**
|
||
* LLM 客户端类，用于与语言模型交互
|
||
*/
|
||
class TransformationLLMClient {
|
||
private llm: LLMManager;
|
||
private model: LLMModel;
|
||
|
||
constructor(llm: LLMManager, model: LLMModel) {
|
||
this.llm = llm;
|
||
this.model = model;
|
||
}
|
||
|
||
async queryChatModel(messages: RequestMessage[]): Promise<Result<string, Error>> {
|
||
try {
|
||
const stream = await this.llm.streamResponse(
|
||
this.model,
|
||
{
|
||
messages: messages,
|
||
model: this.model.modelId,
|
||
stream: true,
|
||
}
|
||
);
|
||
|
||
let response_content = "";
|
||
for await (const chunk of stream) {
|
||
const content = chunk.choices[0]?.delta?.content ?? '';
|
||
response_content += content;
|
||
}
|
||
return ok(response_content);
|
||
} catch (error) {
|
||
return err(error instanceof Error ? error : new Error(String(error)));
|
||
}
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 文档内容处理类
|
||
*/
|
||
class DocumentProcessor {
|
||
private static readonly DEFAULT_MAX_TOKENS = 12000; // 默认最大 token 数
|
||
private static readonly MIN_CONTENT_LENGTH = 100; // 最小内容长度（字符数）
|
||
|
||
/**
|
||
* 检查和处理文档内容大小
|
||
*/
|
||
static async processContent(content: string, maxTokens: number = this.DEFAULT_MAX_TOKENS): Promise<{
|
||
processedContent: string;
|
||
truncated: boolean;
|
||
originalTokens: number;
|
||
processedTokens: number;
|
||
}> {
|
||
const originalTokens = await tokenCount(content);
|
||
|
||
if (originalTokens <= maxTokens) {
|
||
return {
|
||
processedContent: content,
|
||
truncated: false,
|
||
originalTokens,
|
||
processedTokens: originalTokens
|
||
};
|
||
}
|
||
|
||
// 智能截断：基于 token 数量和内容边界
|
||
// 先按字符比例粗略估算截断位置
|
||
const estimatedCharRatio = content.length / originalTokens;
|
||
const estimatedCharLimit = Math.floor(maxTokens * estimatedCharRatio * 0.9); // 留一些缓冲
|
||
|
||
let truncatedContent = content.substring(0, estimatedCharLimit);
|
||
|
||
// 查找最后一个完整句子的结束位置
|
||
const lastSentenceEnd = Math.max(
|
||
truncatedContent.lastIndexOf('.'),
|
||
truncatedContent.lastIndexOf('!'),
|
||
truncatedContent.lastIndexOf('?'),
|
||
truncatedContent.lastIndexOf('。'),
|
||
truncatedContent.lastIndexOf('！'),
|
||
truncatedContent.lastIndexOf('？')
|
||
);
|
||
|
||
// 查找最后一个段落的结束位置
|
||
const lastParagraphEnd = truncatedContent.lastIndexOf('\n\n');
|
||
|
||
// 选择最合适的截断位置
|
||
const cutoffPosition = Math.max(lastSentenceEnd, lastParagraphEnd);
|
||
|
||
if (cutoffPosition > estimatedCharLimit * 0.8) { // 如果截断位置不会丢失太多内容
|
||
truncatedContent = content.substring(0, cutoffPosition + 1);
|
||
}
|
||
|
||
// 确保截断后的内容不会太短
|
||
if (truncatedContent.length < this.MIN_CONTENT_LENGTH) {
|
||
// 按字符比例回退到安全长度
|
||
const safeCharLimit = Math.max(this.MIN_CONTENT_LENGTH, Math.floor(maxTokens * estimatedCharRatio * 0.8));
|
||
truncatedContent = content.substring(0, Math.min(safeCharLimit, content.length));
|
||
}
|
||
|
||
// 验证最终的 token 数量
|
||
const finalTokens = await tokenCount(truncatedContent);
|
||
|
||
// 如果仍然超过限制，进行更精确的截断
|
||
if (finalTokens > maxTokens) {
|
||
const adjustedRatio = truncatedContent.length / finalTokens;
|
||
const adjustedCharLimit = Math.floor(maxTokens * adjustedRatio);
|
||
truncatedContent = content.substring(0, adjustedCharLimit);
|
||
}
|
||
|
||
const processedTokens = await tokenCount(truncatedContent);
|
||
|
||
return {
|
||
processedContent: truncatedContent,
|
||
truncated: true,
|
||
originalTokens,
|
||
processedTokens
|
||
};
|
||
}
|
||
|
||
/**
|
||
* 验证内容是否适合处理
|
||
*/
|
||
static validateContent(content: string): Result<void, Error> {
|
||
if (!content || content.trim().length === 0) {
|
||
return err(new Error('内容不能为空'));
|
||
}
|
||
|
||
if (content.length < this.MIN_CONTENT_LENGTH) {
|
||
return err(new Error(`内容长度至少需要 ${this.MIN_CONTENT_LENGTH} 个字符`));
|
||
}
|
||
|
||
return ok(undefined);
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 主要的转换执行函数
|
||
*/
|
||
export async function runTransformation(params: TransformationParams): Promise<TransformationResult> {
|
||
const { content, transformationType, settings, model, maxContentTokens } = params;
|
||
|
||
try {
|
||
// 验证内容
|
||
const contentValidation = DocumentProcessor.validateContent(content);
|
||
if (contentValidation.isErr()) {
|
||
return {
|
||
success: false,
|
||
error: contentValidation.error.message
|
||
};
|
||
}
|
||
|
||
// 获取转换配置
|
||
const transformationConfig = TRANSFORMATIONS[transformationType];
|
||
if (!transformationConfig) {
|
||
return {
|
||
success: false,
|
||
error: `不支持的转换类型: ${transformationType}`
|
||
};
|
||
}
|
||
|
||
// 处理文档内容（检查 token 数量并截断）
|
||
const tokenLimit = maxContentTokens || DocumentProcessor['DEFAULT_MAX_TOKENS'];
|
||
const processedDocument = await DocumentProcessor.processContent(content, tokenLimit);
|
||
|
||
// 使用默认模型或传入的模型
|
||
const llmModel: LLMModel = model || {
|
||
provider: settings.applyModelProvider,
|
||
modelId: settings.applyModelId,
|
||
};
|
||
|
||
// 创建 LLM 管理器和客户端
|
||
const llmManager = new LLMManager(settings);
|
||
const client = new TransformationLLMClient(llmManager, llmModel);
|
||
|
||
// 构建请求消息
|
||
const messages: RequestMessage[] = [
|
||
{
|
||
role: 'system',
|
||
content: transformationConfig.prompt
|
||
},
|
||
{
|
||
role: 'user',
|
||
content: processedDocument.processedContent
|
||
}
|
||
];
|
||
|
||
// 调用 LLM 执行转换
|
||
const result = await client.queryChatModel(messages);
|
||
|
||
if (result.isErr()) {
|
||
return {
|
||
success: false,
|
||
error: `LLM 调用失败: ${result.error.message}`,
|
||
truncated: processedDocument.truncated,
|
||
originalTokens: processedDocument.originalTokens,
|
||
processedTokens: processedDocument.processedTokens
|
||
};
|
||
}
|
||
|
||
// 后处理结果
|
||
const processedResult = postProcessResult(result.value, transformationType);
|
||
|
||
return {
|
||
success: true,
|
||
result: processedResult,
|
||
truncated: processedDocument.truncated,
|
||
originalTokens: processedDocument.originalTokens,
|
||
processedTokens: processedDocument.processedTokens
|
||
};
|
||
|
||
} catch (error) {
|
||
return {
|
||
success: false,
|
||
error: `转换过程中出现错误: ${error instanceof Error ? error.message : String(error)}`
|
||
};
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 后处理转换结果
|
||
*/
|
||
function postProcessResult(result: string, transformationType: TransformationType): string {
|
||
let processed = result.trim();
|
||
|
||
// 移除可能的 markdown 代码块标记
|
||
processed = processed.replace(/^```[\w]*\n/, '').replace(/\n```$/, '');
|
||
|
||
// 根据转换类型进行特定的后处理
|
||
switch (transformationType) {
|
||
case TransformationType.KEY_INSIGHTS:
|
||
// 确保 insights 格式正确
|
||
if (!processed.includes('INSIGHTS')) {
|
||
processed = `# INSIGHTS\n\n${processed}`;
|
||
}
|
||
break;
|
||
|
||
case TransformationType.REFLECTIONS:
|
||
// 确保 reflections 格式正确
|
||
if (!processed.includes('REFLECTIONS')) {
|
||
processed = `# REFLECTIONS\n\n${processed}`;
|
||
}
|
||
break;
|
||
|
||
case TransformationType.ANALYZE_PAPER: {
|
||
// 确保论文分析包含所有必需的部分
|
||
const requiredSections = ['PURPOSE', 'CONTRIBUTION', 'KEY FINDINGS', 'IMPLICATIONS', 'LIMITATIONS'];
|
||
const hasAllSections = requiredSections.every(section =>
|
||
processed.toUpperCase().includes(section)
|
||
);
|
||
|
||
if (!hasAllSections) {
|
||
// 如果缺少某些部分，添加提示
|
||
processed += '\n\n*注意：某些分析部分可能不完整，建议重新处理或检查原始内容。*';
|
||
}
|
||
break;
|
||
}
|
||
}
|
||
|
||
return processed;
|
||
}
|
||
|
||
/**
|
||
* 批量执行转换
|
||
*/
|
||
export async function runBatchTransformations(
|
||
content: string,
|
||
transformationTypes: TransformationType[],
|
||
settings: InfioSettings,
|
||
model?: LLMModel
|
||
): Promise<Record<string, TransformationResult>> {
|
||
const results: Record<string, TransformationResult> = {};
|
||
|
||
// 并行执行所有转换
|
||
const promises = transformationTypes.map(async (type) => {
|
||
const result = await runTransformation({
|
||
content,
|
||
transformationType: type,
|
||
settings,
|
||
model
|
||
});
|
||
return { type, result };
|
||
});
|
||
|
||
const completedResults = await Promise.all(promises);
|
||
|
||
for (const { type, result } of completedResults) {
|
||
results[type] = result;
|
||
}
|
||
|
||
return results;
|
||
}
|
||
|
||
/**
|
||
* 获取所有可用的转换类型和描述
|
||
*/
|
||
export function getAvailableTransformations(): Array<{ type: TransformationType, description: string }> {
|
||
return Object.values(TRANSFORMATIONS).map(config => ({
|
||
type: config.type,
|
||
description: config.description
|
||
}));
|
||
}
|