import { llm } from "../llm_client.js";
import { structureParser } from "./structure_parser.js";
export const summarizer = {
definition: {
name: "summarizer",
description: "生成智能论文摘要或方法论概述（支持本地 LLM 加速）",
inputSchema: {
type: "object",
properties: {
text: { type: "string", description: "论文全文或部分文本" },
mode: {
type: "string",
enum: ["summary", "methodology", "layman"],
description: "摘要模式：summary(标准摘要), methodology(方法论概述), layman(通俗解释)",
},
},
required: ["text"],
additionalProperties: false,
},
},
handler: async (input: any) => {
const mode = input.mode || "summary";
const text = input.text || "";
// 1. 尝试结构化切分 (如果是全文)
let sections: Record<string, string> = {};
try {
const structRes = await structureParser.handler({ text });
sections = JSON.parse(structRes.content[0].text);
} catch (e) {
// Fallback if parsing fails or text is short
sections = { "Full": text };
}
// 2. 准备上下文
let context = "";
if (mode === "summary") {
// 优先使用 Abstract + Conclusion + Introduction
context = (sections["Abstract"] || "") + "\n" +
(sections["Introduction"] || "").slice(0, 1000) + "\n" +
(sections["Conclusion"] || "");
if (context.length < 100) context = text.slice(0, 2000); // Fallback
} else if (mode === "methodology") {
// 优先使用 Method
context = sections["Method"] || sections["Proposed Method"] || sections["Methodology"] || "";
if (!context) context = text.slice(0, 3000);
} else {
context = (sections["Abstract"] || "") + "\n" + (sections["Introduction"] || "").slice(0, 1000);
}
// 3. 尝试调用本地 LLM
const isLLMAvailable = await llm.isAvailable();
if (isLLMAvailable) {
let prompt = "";
if (mode === "summary") {
prompt = `Please provide a concise academic summary of the following paper content:\n\n${context.slice(0, 4000)}`;
} else if (mode === "methodology") {
prompt = `Please summarize the key methodology and algorithms used in the following text. Focus on the technical details:\n\n${context.slice(0, 4000)}`;
} else {
prompt = `Explain the following scientific content in simple terms for a non-expert:\n\n${context.slice(0, 4000)}`;
}
const llmRes = await llm.chat(prompt, "You are a helpful research assistant.");
if (llmRes) {
return { content: [{ type: "text", text: `[AI Generated] ${llmRes}` }] };
}
}
// 4. Fallback (Heuristic)
let fallbackSummary = "";
if (mode === "summary") {
fallbackSummary = `[Local Extraction]
**Abstract**: ${sections["Abstract"] || "Not found"}
**Conclusion Preview**: ${(sections["Conclusion"] || "").slice(0, 300)}...
(To get an AI summary, ensure a local LLM like Ollama is running on port 11434)`;
} else if (mode === "methodology") {
fallbackSummary = `[Local Extraction]
**Method Section Preview**:
${context.slice(0, 500)}...
(Key technical terms extracted: ${extractKeyTerms(context).join(", ")})`;
} else {
fallbackSummary = `[Local Extraction]
${(sections["Abstract"] || text).slice(0, 500)}...`;
}
return { content: [{ type: "text", text: fallbackSummary }] };
},
};
function extractKeyTerms(text: string): string[] {
// Simple heuristic to find capitalized phrases or math symbols
const matches = text.match(/\b[A-Z][a-zA-Z]+\b/g) || [];
return Array.from(new Set(matches)).slice(0, 10);
}
