| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426 |
- <!DOCTYPE html>
- <html lang="zh-CN">
- <head>
- <meta charset="UTF-8">
- <meta name="viewport" content="width=device-width, initial-scale=1.0">
- <title>PaperMind · 端到端运行流程详解</title>
- <style>
- :root {
- --bg: #0f1117;
- --bg-elev: #161924;
- --bg-card: #1c2030;
- --bg-code: #0a0c14;
- --border: #2a2f42;
- --text: #e6e9f0;
- --text-dim: #9aa0b0;
- --text-mute: #6b7180;
- --accent: #5b9dff;
- --accent-2: #a78bfa;
- --green: #4ade80;
- --yellow: #fbbf24;
- --red: #f87171;
- --cyan: #22d3ee;
- --pink: #f472b6;
- --orange: #fb923c;
- --step1: #60a5fa;
- --step2: #a78bfa;
- --step3: #22d3ee;
- --step4: #4ade80;
- --step5: #fbbf24;
- --step6: #fb923c;
- --step7: #f472b6;
- }
- * { box-sizing: border-box; margin: 0; padding: 0; }
- html { scroll-behavior: smooth; }
- body {
- font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC", "Microsoft YaHei", sans-serif;
- background: var(--bg);
- color: var(--text);
- line-height: 1.65;
- font-size: 15px;
- }
- /* ===== Layout ===== */
- .layout { display: flex; max-width: 1480px; margin: 0 auto; }
- .sidebar {
- position: sticky; top: 0; height: 100vh; width: 260px; flex-shrink: 0;
- border-right: 1px solid var(--border);
- background: var(--bg-elev);
- overflow-y: auto;
- padding: 24px 16px;
- }
- .sidebar h2 { font-size: 13px; color: var(--text-mute); text-transform: uppercase; letter-spacing: 1px; margin-bottom: 12px; padding: 0 8px; }
- .sidebar a {
- display: block; padding: 7px 12px; color: var(--text-dim); text-decoration: none;
- border-radius: 6px; font-size: 13.5px; transition: all 0.15s;
- border-left: 2px solid transparent;
- }
- .sidebar a:hover { color: var(--text); background: rgba(91,157,255,0.08); }
- .sidebar a.active { color: var(--accent); background: rgba(91,157,255,0.12); border-left-color: var(--accent); }
- .sidebar a.sub { padding-left: 24px; font-size: 12.5px; }
- .main { flex: 1; min-width: 0; padding: 48px 56px 120px; }
- /* ===== Hero ===== */
- .hero {
- background: linear-gradient(135deg, #1a1f3a 0%, #0f1117 60%);
- border: 1px solid var(--border);
- border-radius: 16px;
- padding: 48px 44px;
- margin-bottom: 48px;
- position: relative;
- overflow: hidden;
- }
- .hero::before {
- content: ""; position: absolute; top: -50%; right: -10%; width: 500px; height: 500px;
- background: radial-gradient(circle, rgba(91,157,255,0.15), transparent 70%);
- pointer-events: none;
- }
- .hero h1 { font-size: 34px; font-weight: 700; margin-bottom: 12px; background: linear-gradient(90deg, #fff, #9aa0b0); -webkit-background-clip: text; -webkit-text-fill-color: transparent; }
- .hero .sub { font-size: 16px; color: var(--text-dim); max-width: 760px; }
- .hero .meta { margin-top: 24px; display: flex; gap: 12px; flex-wrap: wrap; }
- .badge { background: rgba(91,157,255,0.12); border: 1px solid rgba(91,157,255,0.3); color: var(--accent); padding: 4px 12px; border-radius: 999px; font-size: 12.5px; }
- /* ===== Section ===== */
- section { margin-bottom: 56px; scroll-margin-top: 20px; }
- h2.section-title {
- font-size: 24px; font-weight: 700; margin-bottom: 8px;
- padding-bottom: 12px; border-bottom: 1px solid var(--border);
- display: flex; align-items: center; gap: 10px;
- }
- h2.section-title .num {
- display: inline-flex; align-items: center; justify-content: center;
- width: 32px; height: 32px; border-radius: 8px;
- background: linear-gradient(135deg, var(--accent), var(--accent-2));
- color: #fff; font-size: 15px; font-weight: 700;
- }
- .section-desc { color: var(--text-dim); margin-bottom: 24px; font-size: 14.5px; }
- h3 { font-size: 18px; font-weight: 600; margin: 28px 0 14px; color: var(--text); }
- h4 { font-size: 15px; font-weight: 600; margin: 18px 0 10px; color: var(--accent); }
- p { margin-bottom: 10px; }
- /* ===== Cards ===== */
- .card {
- background: var(--bg-card);
- border: 1px solid var(--border);
- border-radius: 12px;
- padding: 22px 24px;
- margin-bottom: 18px;
- }
- .card.glow { box-shadow: 0 0 0 1px rgba(91,157,255,0.15), 0 8px 32px rgba(0,0,0,0.3); }
- .step-card { border-left: 3px solid var(--accent); }
- .step-card.s1 { border-left-color: var(--step1); }
- .step-card.s2 { border-left-color: var(--step2); }
- .step-card.s3 { border-left-color: var(--step3); }
- .step-card.s4 { border-left-color: var(--step4); }
- .step-card.s5 { border-left-color: var(--step5); }
- .step-card.s6 { border-left-color: var(--step6); }
- .step-card.s7 { border-left-color: var(--step7); }
- .step-header { display: flex; align-items: center; justify-content: space-between; margin-bottom: 14px; flex-wrap: wrap; gap: 10px; }
- .step-tag {
- display: inline-flex; align-items: center; gap: 8px;
- padding: 5px 12px; border-radius: 6px; font-size: 13px; font-weight: 600;
- }
- .step-tag.s1 { background: rgba(96,165,250,0.15); color: var(--step1); }
- .step-tag.s2 { background: rgba(167,139,250,0.15); color: var(--step2); }
- .step-tag.s3 { background: rgba(34,211,238,0.15); color: var(--step3); }
- .step-tag.s4 { background: rgba(74,222,128,0.15); color: var(--step4); }
- .step-tag.s5 { background: rgba(251,191,36,0.15); color: var(--step5); }
- .step-tag.s6 { background: rgba(251,146,60,0.15); color: var(--step6); }
- .step-tag.s7 { background: rgba(244,114,182,0.15); color: var(--step7); }
- .step-num { font-size: 18px; font-weight: 700; }
- .step-title { font-size: 17px; font-weight: 600; }
- .code-link {
- font-family: "JetBrains Mono", Consolas, monospace; font-size: 12px;
- color: var(--text-dim); text-decoration: none;
- background: rgba(0,0,0,0.3); padding: 3px 8px; border-radius: 4px;
- border: 1px solid var(--border);
- }
- .code-link:hover { color: var(--accent); border-color: var(--accent); }
- /* ===== IO blocks ===== */
- .io-grid { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; margin: 14px 0; }
- @media (max-width: 900px) { .io-grid { grid-template-columns: 1fr; } }
- .io-box {
- background: var(--bg-code); border: 1px solid var(--border); border-radius: 8px;
- padding: 14px 16px;
- }
- .io-box .label {
- font-size: 11px; font-weight: 700; text-transform: uppercase; letter-spacing: 1px;
- margin-bottom: 8px; display: flex; align-items: center; gap: 6px;
- }
- .io-box.in .label { color: var(--cyan); }
- .io-box.out .label { color: var(--green); }
- .io-box .label::before { content: ""; width: 8px; height: 8px; border-radius: 50%; }
- .io-box.in .label::before { background: var(--cyan); }
- .io-box.out .label::before { background: var(--green); }
- /* ===== Code blocks ===== */
- pre {
- background: var(--bg-code); border: 1px solid var(--border); border-radius: 8px;
- padding: 14px 16px; overflow-x: auto; font-size: 12.5px; line-height: 1.55;
- font-family: "JetBrains Mono", Consolas, "Courier New", monospace;
- color: #d4d4dc;
- }
- pre.prompt {
- background: linear-gradient(180deg, #14111f 0%, #0a0c14 100%);
- border-color: rgba(167,139,250,0.3);
- }
- code.inline {
- background: rgba(91,157,255,0.12); color: var(--accent);
- padding: 1px 6px; border-radius: 4px; font-size: 12.5px;
- font-family: "JetBrains Mono", Consolas, monospace;
- }
- .tok-sys { color: #c792ea; }
- .tok-user { color: #82aaff; }
- .tok-kw { color: #c792ea; }
- .tok-str { color: #c3e88d; }
- .tok-com { color: #6b7180; font-style: italic; }
- .tok-num { color: #f78c6c; }
- /* ===== Collapsible ===== */
- details {
- background: var(--bg-code); border: 1px solid var(--border); border-radius: 8px;
- margin: 10px 0; overflow: hidden;
- }
- details[open] { border-color: rgba(167,139,250,0.4); }
- summary {
- cursor: pointer; padding: 12px 16px; font-size: 13.5px; font-weight: 600;
- color: var(--accent-2); user-select: none;
- display: flex; align-items: center; gap: 8px;
- background: linear-gradient(90deg, rgba(167,139,250,0.08), transparent);
- }
- summary:hover { background: rgba(167,139,250,0.1); }
- summary::before { content: "▸"; transition: transform 0.15s; font-size: 11px; }
- details[open] summary::before { transform: rotate(90deg); }
- details > pre, details > div { border: none; border-radius: 0; margin: 0; border-top: 1px solid var(--border); }
- /* ===== Tables ===== */
- table {
- width: 100%; border-collapse: collapse; margin: 14px 0;
- background: var(--bg-card); border: 1px solid var(--border); border-radius: 8px;
- overflow: hidden; font-size: 13px;
- }
- th, td { padding: 10px 14px; text-align: left; border-bottom: 1px solid var(--border); vertical-align: top; }
- th { background: rgba(91,157,255,0.06); color: var(--accent); font-weight: 600; font-size: 12px; text-transform: uppercase; letter-spacing: 0.5px; }
- tr:last-child td { border-bottom: none; }
- tr:hover td { background: rgba(91,157,255,0.04); }
- /* ===== Pipeline diagram ===== */
- .pipeline {
- display: flex; flex-direction: column; gap: 0; margin: 24px 0;
- }
- .pipeline-row {
- display: flex; align-items: stretch; gap: 0;
- }
- .pipe-node {
- flex: 1; background: var(--bg-card); border: 1px solid var(--border);
- border-radius: 10px; padding: 14px; position: relative;
- transition: transform 0.15s;
- }
- .pipe-node:hover { transform: translateY(-2px); }
- .pipe-node .pn-num {
- font-size: 11px; font-weight: 700; color: var(--text-mute); margin-bottom: 4px;
- }
- .pipe-node .pn-title { font-size: 14px; font-weight: 600; margin-bottom: 4px; }
- .pipe-node .pn-desc { font-size: 11.5px; color: var(--text-dim); line-height: 1.5; }
- .pipe-node .pn-llm { font-size: 10.5px; margin-top: 6px; color: var(--accent-2); }
- .pipe-arrow {
- display: flex; align-items: center; justify-content: center;
- width: 28px; color: var(--text-mute); font-size: 18px;
- }
- /* ===== Tags ===== */
- .tag {
- display: inline-block; padding: 2px 8px; border-radius: 4px;
- font-size: 11px; font-weight: 600; margin-right: 4px;
- }
- .tag.llm { background: rgba(167,139,250,0.15); color: var(--accent-2); }
- .tag.tool { background: rgba(34,211,238,0.15); color: var(--cyan); }
- .tag.local { background: rgba(74,222,128,0.15); color: var(--green); }
- .tag.cache { background: rgba(251,191,36,0.15); color: var(--yellow); }
- .tag.degrade { background: rgba(248,113,113,0.15); color: var(--red); }
- .tag.parallel { background: rgba(244,114,182,0.15); color: var(--pink); }
- /* ===== List ===== */
- ul, ol { margin: 8px 0 12px 22px; }
- li { margin-bottom: 5px; }
- li code.inline { font-size: 12px; }
- .callout {
- background: rgba(91,157,255,0.06); border-left: 3px solid var(--accent);
- padding: 12px 16px; border-radius: 0 8px 8px 0; margin: 12px 0; font-size: 13.5px;
- }
- .callout.warn { background: rgba(251,191,36,0.06); border-color: var(--yellow); }
- .callout.danger { background: rgba(248,113,113,0.06); border-color: var(--red); }
- .callout.ok { background: rgba(74,222,128,0.06); border-color: var(--green); }
- .flow-arrow { color: var(--text-mute); text-align: center; font-size: 18px; margin: 6px 0; }
- .pill-row { display: flex; flex-wrap: wrap; gap: 6px; margin: 8px 0; }
- .keyword { color: #c3e88d; }
- .placeholder { color: #f78c6c; }
- .footer { text-align: center; color: var(--text-mute); padding: 40px 0 0; border-top: 1px solid var(--border); margin-top: 60px; font-size: 12.5px; }
- ::-webkit-scrollbar { width: 10px; height: 10px; }
- ::-webkit-scrollbar-track { background: var(--bg); }
- ::-webkit-scrollbar-thumb { background: var(--border); border-radius: 5px; }
- ::-webkit-scrollbar-thumb:hover { background: #3a4060; }
- </style>
- </head>
- <body>
- <div class="layout">
- <!-- Sidebar -->
- <nav class="sidebar">
- <h2>导航</h2>
- <a href="#overview">总览</a>
- <a href="#startup">1. 启动阶段</a>
- <a href="#submit">2. 提交任务</a>
- <a href="#phase1">3. Phase 1 预处理</a>
- <a href="#step1" class="sub">① PDF 解析</a>
- <a href="#step2" class="sub">② 结构化抽取</a>
- <a href="#step3" class="sub">③ 向量索引</a>
- <a href="#phase2">4. Phase 2 Plan</a>
- <a href="#phase3">5. Phase 3 Execute</a>
- <a href="#step4" class="sub">④ Gap 识别</a>
- <a href="#reflexion" class="sub">↳ Reflexion</a>
- <a href="#step5" class="sub">⑤ 创新点生成</a>
- <a href="#step6" class="sub">⑥ 实验设计</a>
- <a href="#step7" class="sub">⑦ 整合输出</a>
- <a href="#eval">6. 评估阶段</a>
- <a href="#summary">7. LLM 调用清单</a>
- <a href="#principles">8. Prompt 设计原则</a>
- <a href="#degrade">9. 降级与容错</a>
- </nav>
- <!-- Main -->
- <main class="main">
- <!-- Hero -->
- <div class="hero">
- <h1>PaperMind · 端到端运行流程详解</h1>
- <p class="sub">从用户提交 PDF 文件夹,到生成结构化研究提案(含创新点 + 实验方案)的完整代码执行路径、每次 LLM 调用与工具调用的输入输出全记录。</p>
- <div class="meta">
- <span class="badge">Plan-and-Execute + Reflexion</span>
- <span class="badge">7 步漏斗式信息收敛</span>
- <span class="badge">DeepSeek Function Calling</span>
- <span class="badge">BGE-m3 + Chroma 混合检索</span>
- </div>
- </div>
- <!-- Overview -->
- <section id="overview">
- <h2 class="section-title"><span class="num">0</span>总览:七步漏斗流水线</h2>
- <p class="section-desc">整个系统由 <code class="inline">WorkflowOrchestrator</code> 编排,分三大阶段:预处理(步骤 1-3,带缓存)→ Plan(LLM 动态规划步骤 4-7)→ Execute(步骤 4-7,每步后插入 Reflexion 自校验)。</p>
- <div class="pipeline">
- <div class="pipeline-row">
- <div class="pipe-node" style="border-left:3px solid var(--step1)">
- <div class="pn-num">STEP 1</div>
- <div class="pn-title">PDF 解析</div>
- <div class="pn-desc">PyMuPDF 提取章节结构</div>
- <div class="pn-llm"><span class="tag local">本地</span></div>
- </div>
- <div class="pipe-arrow">→</div>
- <div class="pipe-node" style="border-left:3px solid var(--step2)">
- <div class="pn-num">STEP 2</div>
- <div class="pn-title">结构化抽取</div>
- <div class="pn-desc">Function Calling 抽 9 字段</div>
- <div class="pn-llm"><span class="tag llm">LLM</span> <span class="tag parallel">并行</span></div>
- </div>
- <div class="pipe-arrow">→</div>
- <div class="pipe-node" style="border-left:3px solid var(--step3)">
- <div class="pn-num">STEP 3</div>
- <div class="pn-title">向量索引</div>
- <div class="pn-desc">BGE-m3 → Chroma + BM25</div>
- <div class="pn-llm"><span class="tag local">本地</span></div>
- </div>
- </div>
- <div class="flow-arrow">↓</div>
- <div class="pipeline-row">
- <div class="pipe-node" style="border-left:3px solid var(--step4)">
- <div class="pn-num">STEP 4</div>
- <div class="pn-title">Gap 识别</div>
- <div class="pn-desc">四类 Gap 信号综合推理</div>
- <div class="pn-llm"><span class="tag llm">LLM</span> <span class="tag tool">RAG</span></div>
- </div>
- <div class="pipe-arrow">→</div>
- <div class="pipe-node" style="border-left:3px solid var(--step5)">
- <div class="pn-num">STEP 5</div>
- <div class="pn-title">创新点生成</div>
- <div class="pn-desc">四类来源 + 三维评分筛选</div>
- <div class="pn-llm"><span class="tag llm">LLM</span> <span class="tag tool">联网</span></div>
- </div>
- <div class="pipe-arrow">→</div>
- <div class="pipe-node" style="border-left:3px solid var(--step6)">
- <div class="pn-num">STEP 6</div>
- <div class="pn-title">实验设计</div>
- <div class="pn-desc">8 要素可执行方案</div>
- <div class="pn-llm"><span class="tag llm">LLM</span></div>
- </div>
- <div class="pipe-arrow">→</div>
- <div class="pipe-node" style="border-left:3px solid var(--step7)">
- <div class="pn-num">STEP 7</div>
- <div class="pn-title">整合输出</div>
- <div class="pn-desc">Markdown 研究提案</div>
- <div class="pn-llm"><span class="tag llm">LLM</span></div>
- </div>
- </div>
- </div>
- <div class="callout">
- <strong>核心设计哲学:</strong>漏斗式信息收敛 — 每步压缩信息,最后一步才做创造性推理(避免早期噪声干扰)。所有中间产物都是结构化 JSON / 向量,不依赖长上下文记忆。
- </div>
- </section>
- <!-- Startup -->
- <section id="startup">
- <h2 class="section-title"><span class="num">1</span>启动阶段(进程级)</h2>
- <p class="section-desc">用户运行 <code class="inline">python run_backend.py</code> 启动 FastAPI 服务,前端 <code class="inline">npm run dev</code> 启动 Vite。</p>
- <div class="card">
- <h4>启动命令</h4>
- <pre>python run_backend.py
- # 等价于:uvicorn backend.api.main:app --host 0.0.0.0 --port 8000</pre>
- <h4>启动钩子(on_startup)</h4>
- <p>应用启动时预加载重型依赖,避免首次请求时才触发 import(torch 单独 import 100+ 秒):</p>
- <pre><span class="tok-com"># 顺序预加载</span>
- import torch <span class="tok-com"># ~5s</span>
- from transformers import AutoModel, AutoTokenizer <span class="tok-com"># ~3s</span>
- import chromadb <span class="tok-com"># ~2s</span>
- <span class="tok-com"># 预热 Embedder 单例:加载 BGE-m3 模型 + 首次 forward(初始化 CUDA 上下文)</span>
- _embedder = get_default_embedder()
- _embedder.embed_texts([<span class="tok-str">"预加载测试文本"</span>])</pre>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <div>无(进程启动)</div>
- </div>
- <div class="io-box out">
- <div class="label">输出</div>
- <div>FastAPI 服务监听 <code class="inline">0.0.0.0:8000</code>,BGE 模型已 warmup,前端可访问 <code class="inline">localhost:5173</code></div>
- </div>
- </div>
- </div>
- </section>
- <!-- Submit -->
- <section id="submit">
- <h2 class="section-title"><span class="num">2</span>用户提交分析任务</h2>
- <p class="section-desc">前端发 HTTP 请求,后端创建任务并后台异步启动,立即返回 task_id。</p>
- <div class="card">
- <h4>HTTP 请求</h4>
- <pre><span class="tok-com"># POST /api/analyze</span>
- {
- <span class="tok-str">"folder_path"</span>: <span class="tok-str">"D:/papers"</span>,
- <span class="tok-str">"research_direction"</span>: <span class="tok-str">"多模态学习"</span>
- }</pre>
- <h4>路由处理 <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/api/routes.py#L157">routes.py:157</a></h4>
- <ol>
- <li><code class="inline">task_manager.create_task()</code> → 生成 <code class="inline">task_id</code>(uuid),状态 <code class="inline">pending</code></li>
- <li><code class="inline">asyncio.create_task(run_analysis(...))</code> → 后台异步启动,<strong>不阻塞响应</strong></li>
- <li>立即返回 <code class="inline">{"task_id": "..."}</code></li>
- </ol>
- <h4>后台主流程 <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/api/routes.py#L51">routes.py:51</a></h4>
- <pre><span class="tok-com"># 初始化组件</span>
- client = DeepSeekClient() <span class="tok-com"># LLM 客户端</span>
- registry = build_default_registry() <span class="tok-com"># 工具注册表</span>
- memory = MemoryManager() <span class="tok-com"># 双层记忆</span>
- orchestrator = WorkflowOrchestrator(client, registry, memory)
- <span class="tok-com"># 执行流水线</span>
- pipeline_output = await orchestrator.run_pipeline(...)
- <span class="tok-com"># 生成报告 + 评估</span>
- report = await ReportAgent(client).generate_report(final_results)
- evaluation = await Evaluator(client).evaluate(...)
- <span class="tok-com"># 存结果</span>
- task_manager.update_task(task_id, status=<span class="tok-str">"completed"</span>, result={...})</pre>
- <div class="callout ok">
- <strong>SSE 进度推送:</strong>整个后台流程通过 <code class="inline">progress_callback</code> 实时推送 <code class="inline">{stage, status, message, progress, payload}</code> 到 <code class="inline">GET /api/progress/{task_id}</code>,前端可流式渲染进度条。
- </div>
- </div>
- </section>
- <!-- Phase 1 -->
- <section id="phase1">
- <h2 class="section-title"><span class="num">3</span>Phase 1:预处理(步骤 1-3,带缓存)</h2>
- <p class="section-desc">由 <code class="inline">Controller._preprocess_papers()</code> 编排,对每篇论文检查 PaperLibrary 缓存状态:已完成的步骤直接从磁盘加载(标记 cached),未完成的执行后写盘。</p>
- </section>
- <!-- Step 1 -->
- <section id="step1">
- <div class="card step-card s1">
- <div class="step-header">
- <div>
- <span class="step-tag s1"><span class="step-num">①</span> PDF 解析</span>
- <span class="step-title">PyMuPDF 提取章节结构</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/tools/parse_pdf.py">parse_pdf.py</a>
- </div>
- <div class="pill-row">
- <span class="tag local">本地处理</span>
- <span class="tag parallel">并发 asyncio.gather</span>
- <span class="tag cache">缓存 data/cache/parsed/{paper_id}.json</span>
- </div>
- <h4>处理流程</h4>
- <ul>
- <li>PyMuPDF 提取每页文本 + 字号/字体信息</li>
- <li>启发式识别章节标题(字号 > 正文×1.15 或加粗 + 关键词匹配 <code class="inline">abstract|introduction|method...</code>)</li>
- <li>去页眉页脚(短行 < 80 字符且重复出现)</li>
- <li>提取图表标题(<code class="inline">Figure X</code> / <code class="inline">表 X</code> 正则)</li>
- <li>合并跨页段落,修复断句</li>
- </ul>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- <span class="tok-str">"file_path"</span>: <span class="tok-str">"D:/papers/xxx.pdf"</span>
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出:ParsedPaper</div>
- <pre>{
- <span class="tok-str">"path"</span>: <span class="tok-str">"D:/papers/xxx.pdf"</span>,
- <span class="tok-str">"title"</span>: <span class="tok-str">"Attention Is All You Need"</span>,
- <span class="tok-str">"sections"</span>: [
- {
- <span class="tok-str">"heading"</span>: <span class="tok-str">"Abstract"</span>,
- <span class="tok-str">"content"</span>: <span class="tok-str">"..."</span>,
- <span class="tok-str">"figure_captions"</span>: [<span class="tok-str">"Figure 1: ..."</span>]
- }
- ]
- }</pre>
- </div>
- </div>
- <div class="callout">
- <strong>缓存复用:</strong>解析结果写入 <code class="inline">data/cache/parsed/{paper_id}.json</code>,PaperLibrary 标记 <code class="inline">parse_status=done</code>。下次同一论文直接 <code class="inline">load_parsed()</code> 加载,跳过解析。
- </div>
- </div>
- </section>
- <!-- Step 2 -->
- <section id="step2">
- <div class="card step-card s2">
- <div class="step-header">
- <div>
- <span class="step-tag s2"><span class="step-num">②</span> 结构化抽取</span>
- <span class="step-title">核心主干 · Function Calling 强制 JSON</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/tools/extract_paper_structure.py#L214">extract_paper_structure.py:214</a>
- </div>
- <div class="pill-row">
- <span class="tag llm">LLM 调用</span>
- <span class="tag parallel">并行 asyncio.gather + Semaphore(5)</span>
- <span class="tag cache">缓存 data/cache/papers/{paper_id}.json + papers.jsonl</span>
- </div>
- <h4>处理流程</h4>
- <ol>
- <li><strong>长度判断</strong>:总字符 > 20000 → 按章节分块抽取后合并;否则一次性抽取</li>
- <li><strong>调用 LLM</strong>(<code class="inline">_call_extract</code>):强制 <code class="inline">tool_choice={"type":"function","function":{"name":"extract_paper_fields"}}</code></li>
- <li><strong>字段完整性校验</strong>:缺失字段 → 针对性补抽(<code class="inline">focus_fields</code> 只补缺失的,不重跑整篇)</li>
- <li><strong>长论文合并</strong>:多块记录按置信度取最高 value,<code class="inline">source_sections</code> 去重合并</li>
- </ol>
- <details open>
- <summary>LLM Prompt · 抽取(核心)</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是论文结构化抽取助手。严格通过函数调用输出 JSON,
- 每个字段含 value/source_sections/confidence。
- limitations 必须真实可溯源,严禁臆造。
- <span class="tok-user">[User]</span>
- 请从以下论文中抽取结构化字段。
- 论文标题:<span class="placeholder">{title}</span>
- 论文正文:
- ## Abstract
- <span class="placeholder">{...}</span>
- ## Introduction
- <span class="placeholder">{...}</span>
- 抽取要求:
- 1. 必须抽取以下字段:title、task_problem、method、key_contributions、
- datasets、metrics、results、limitations、future_work
- 2. 每个字段必须是 FieldSource 结构:
- {
- "value": "字段值(字符串,尽量简洁完整)",
- "source_sections": ["原文出处章节标题列表"],
- "confidence": 0.0-1.0
- }
- 3. source_sections 必须标注字段信息来源于上面哪个章节(用章节标题原文),便于回溯核查
- 4. <span class="tok-kw">limitations 字段特别要求</span>:必须是论文明确承认的局限,或可从实验结果直接推断的局限,
- 严禁臆造或泛泛而谈。若论文未明确提及局限,value 填写"论文未明确陈述局限",confidence 设为较低值(如 0.3)
- 5. confidence 反映可信度:
- - 信息明确且原文直述 → 0.9-1.0
- - 需要从原文推断 → 0.6-0.8
- - 信息不足或模糊 → 低于 0.6
- 6. 严格输出 JSON,不要添加任何额外说明文字或 markdown 代码块标记</pre>
- </details>
- <details>
- <summary>LLM Prompt · 针对性补抽(缺失字段)</summary>
- <pre class="prompt"><span class="tok-user">[User]</span>
- 前一次抽取缺失了以下字段:<span class="placeholder">{", ".join(focus_fields)}</span>。
- 请重点从原文中补抽这些字段,其他字段可留空。
- <span class="tok-com">{...同上 prompt...}</span></pre>
- </details>
- <details>
- <summary>Function Calling Schema · extract_paper_fields</summary>
- <pre>{
- "type": "function",
- "function": {
- "name": "extract_paper_fields",
- "description": "从论文中抽取结构化字段,每个字段含 value/source_sections/confidence",
- "parameters": {
- "type": "object",
- "properties": {
- "title": { "type": "object", "properties": {
- "value": {"type":"string"},
- "source_sections": {"type":"array","items":{"type":"string"}},
- "confidence": {"type":"number"}
- }},
- "task_problem": { ... 同上 ... },
- "method": { ... },
- "key_contributions": { ... },
- "datasets": { ... },
- "metrics": { ... },
- "results": { ... },
- "limitations": { ... },
- "future_work": { ... }
- },
- "required": ["title","task_problem","method","key_contributions",
- "datasets","metrics","results","limitations","future_work"]
- }
- }
- }</pre>
- </details>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- <span class="tok-str">"paper_path"</span>: <span class="tok-str">"D:/papers/xxx.pdf"</span>,
- <span class="tok-str">"parsed_paper"</span>: <span class="tok-com">{ParsedPaper dict}</span>
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出:PaperRecord</div>
- <pre>{
- "path": "...",
- "title": {
- "value": "Attention Is All You Need",
- "source_sections": ["Title"],
- "confidence": 1.0
- },
- "task_problem": {
- "value": "序列建模任务...",
- "source_sections": ["Introduction"],
- "confidence": 0.9
- },
- "method": {
- "value": "Transformer 架构...",
- "source_sections": ["3 Method","3.1"],
- "confidence": 0.95
- },
- "limitations": {
- "value": "位置编码外推性差...",
- "source_sections": ["6 Conclusion"],
- "confidence": 0.7
- },
- <span class="tok-com">...</span>
- }</pre>
- </div>
- </div>
- <div class="callout warn">
- <strong>防编造机制:</strong>每个字段强制 <code class="inline">source_sections</code> 回溯到具体章节,<code class="inline">limitations</code> 严禁臆造。这是后续创新点论证可追溯性的基石。
- </div>
- </div>
- </section>
- <!-- Step 3 -->
- <section id="step3">
- <div class="card step-card s3">
- <div class="step-header">
- <div>
- <span class="step-tag s3"><span class="step-num">③</span> 向量索引构建</span>
- <span class="step-title">BGE-m3 → Chroma + BM25</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/tools/build_vector_index.py">build_vector_index.py</a>
- </div>
- <div class="pill-row">
- <span class="tag local">本地处理</span>
- <span class="tag cache">Chroma PersistentClient 持久化</span>
- <span class="tag degrade">失败降级 skip_rag</span>
- </div>
- <h4>处理流程</h4>
- <ol>
- <li><strong>分块</strong>:按段落累积到 ~500 字符落块,每块携带 <code class="inline">section</code> 标题</li>
- <li><strong>向量化</strong>:BGE-m3 本地 Embedding(无 API 成本),块内容 = <code class="inline">section_heading + text</code></li>
- <li><strong>写入 Chroma</strong>:<code class="inline">paper_chunks</code> collection,metadata 含 <code class="inline">paper_path/title/section</code></li>
- <li><strong>同步构建 BM25</strong>:倒排索引供关键词检索</li>
- </ol>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- <span class="tok-str">"papers_json"</span>: [<span class="tok-com">{ParsedPaper}</span>, ...]
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出</div>
- <pre>{
- "indexed_count": 12,
- "status": "done"
- }</pre>
- </div>
- </div>
- <div class="callout">
- <strong>定位:</strong>RAG 是"点查询工具"而非"分析引擎",仅在结构化摘要不够用时调用(步骤④⑤ 中需原文细节佐证时)。混合搜索 = Chroma 向量(α=0.6)+ BM25(α=0.4)分数归一化加权融合。
- </div>
- </div>
- </section>
- <!-- Phase 2 -->
- <section id="phase2">
- <h2 class="section-title"><span class="num">4</span>Phase 2:Plan(LLM 动态规划)</h2>
- <p class="section-desc">让 LLM 根据任务自主规划步骤顺序/增减,体现动态编排。失败时回退默认七步计划。实际执行时过滤掉步骤 1-3(已由预处理完成),只执行步骤 4-7。</p>
- <div class="card">
- <h4>调用位置 <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/orchestrator/controller.py#L126">controller.py:126</a></h4>
- <details open>
- <summary>LLM Prompt · 动态规划</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是 Agent 编排器,负责生成可执行的动态计划。只输出 JSON。
- <span class="tok-user">[User]</span>
- 你是论文创新 Agent 的 Controller。请为以下任务生成动态执行计划。
- 任务:分析文件夹 <span class="placeholder">{folder_path}</span> 中的 PDF 论文,
- 识别研究空白,生成 2-3 个创新点并设计实验方案。
- 可用 Agent:
- - PaperParser:扫描文件夹并解析 PDF
- - Extractor:单篇论文结构化抽取
- - Indexer:向量索引构建
- - GapAnalyzer:跨论文 Gap 识别
- - InnovationGenerator:创新点生成与筛选
- - ExperimentDesigner:实验方案设计
- - Integration:整合输出研究提案
- 请生成执行计划,输出 JSON 数组,每个元素含:
- - step: 步骤序号(整数)
- - agent: Agent 名称(上述之一)
- - action: 该步骤具体任务描述
- - depends_on: 依赖的步骤序号列表
- 你可以根据任务自主调整顺序或增减步骤。
- 默认链路:解析→抽取→索引→Gap分析→创新生成→实验设计→整合。
- 严格只输出 JSON 数组,不要其他文本。</pre>
- </details>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre><span class="tok-str">"D:/papers"</span> <span class="tok-com">或</span>
- <span class="tok-str">"已上传的 12 篇论文"</span></pre>
- </div>
- <div class="io-box out">
- <div class="label">输出:计划列表</div>
- <pre>[
- {"step":4,"agent":"GapAnalyzer",
- "action":"跨论文综合分析识别研究空白",
- "depends_on":[2]},
- {"step":5,"agent":"InnovationGenerator",
- "action":"针对 Gap 生成创新点并筛选",
- "depends_on":[4]},
- <span class="tok-com">...</span>
- ]</pre>
- </div>
- </div>
- </div>
- </section>
- <!-- Phase 3 -->
- <section id="phase3">
- <h2 class="section-title"><span class="num">5</span>Phase 3:Execute(步骤 4-7 + Reflexion)</h2>
- <p class="section-desc">逐步执行计划,每步完成后插入 Reflexion 自校验钩子(仅对 Extractor / GapAnalyzer / InnovationGenerator 启用)。每步还根据结果动态调整剩余计划。</p>
- <div class="callout warn">
- <strong>绕过 Function Calling 的设计:</strong>步骤 ④⑤⑥ 中,Controller 直接调用工具(参数已确定),不让模型再选一次工具。因为 DeepSeek 实测会把整段 JSON 当字符串传或漏传参数,浪费迭代次数。只让 LLM 做推理,不让它做参数路由。
- </div>
- </section>
- <!-- Step 4 -->
- <section id="step4">
- <div class="card step-card s4">
- <div class="step-header">
- <div>
- <span class="step-tag s4"><span class="step-num">④</span> Gap 识别</span>
- <span class="step-title">跨论文综合推理 · 四类 Gap 信号</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/tools/analyze_gaps.py#L88">analyze_gaps.py:88</a>
- </div>
- <div class="pill-row">
- <span class="tag llm">LLM 调用</span>
- <span class="tag tool">hybrid_search 工具</span>
- <span class="tag parallel">RAG 增强 evidence</span>
- </div>
- <h4>处理流程</h4>
- <ol>
- <li><strong>聚合对比矩阵</strong>:把 N 篇论文的 8 个字段聚合成文本(关键:让模型在全局视野下识别 Gap)</li>
- <li><strong>调用 LLM</strong> 做跨论文综合推理,识别四类 Gap</li>
- <li><strong>RAG 增强</strong>:对每个 Gap 调用 <code class="inline">hybrid_search</code> 检索原文细节,补充 evidence</li>
- </ol>
- <details open>
- <summary>LLM Prompt · Gap 识别</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是一位严谨的学术研究分析专家,擅长跨论文综合推理与 Gap 识别。
- 只输出严格 JSON,不要输出任何其他内容。
- <span class="tok-user">[User]</span>
- 你是一位资深学术研究者,需要从以下多篇论文的结构化对比中识别研究空白(Gap)。
- ## 论文对比矩阵
- ===== 论文 1: Attention Is All You Need =====
- [研究任务/问题] 序列建模任务...
- [核心方法] Transformer 架构...
- [局限性] 位置编码外推性差...
- [未来工作] 扩展到多模态...
- ===== 论文 2: BERT =====
- [研究任务/问题] ...
- [核心方法] ...
- [局限性] ...
- ## 任务
- 请综合分析上述论文,识别以下四类 Gap 信号:
- 1. <span class="tok-kw">repeated_limitation</span>(重复局限):多篇论文提到同一局限性,但无人解决
- 2. <span class="tok-kw">method_gap</span>(方法空白):某类问题只有少数方法尝试且效果不佳
- 3. <span class="tok-kw">contradictory</span>(矛盾结论):不同论文对同一问题给出冲突结论
- 4. <span class="tok-kw">unfulfilled_future</span>(未兑现的未来工作):多篇论文指向同一未来方向但无人完成
- 对于单篇论文场景,重点关注该论文自身的 limitations 与 future_work 字段。
- ## 输出要求
- 严格输出 JSON:
- {
- "gaps": [
- {
- "description": "Gap 的清晰描述",
- "gap_type": "repeated_limitation|method_gap|contradictory|unfulfilled_future",
- "source_papers": ["来源论文标题1"],
- "evidence": "论证依据,引用具体论文的具体字段",
- "confidence": 0.0到1.0之间的置信度
- }
- ]
- }
- 只输出真正有证据支撑的 Gap,不要编造。每个 Gap 的 source_papers 至少包含 2 篇论文。</pre>
- </details>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- <span class="tok-str">"papers_records"</span>: [
- <span class="tok-com">{PaperRecord 1}</span>,
- <span class="tok-com">{PaperRecord 2}</span>,
- ...
- ]
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出</div>
- <pre>{
- "gaps": [
- {
- "description": "位置编码外推性普遍较差",
- "gap_type": "repeated_limitation",
- "source_papers": ["Attention...","BERT..."],
- "evidence": "两篇均提及... [原文佐证]\n - 《Attention》6 Conclusion: ...",
- "confidence": 0.85
- }
- ]
- }</pre>
- </div>
- </div>
- </div>
- </section>
- <!-- Reflexion -->
- <section id="reflexion">
- <div class="card step-card" style="border-left-color: var(--accent-2);">
- <div class="step-header">
- <div>
- <span class="step-tag" style="background:rgba(167,139,250,0.15);color:var(--accent-2);">↻ Reflexion 自校验</span>
- <span class="step-title">针对性补抽(非整篇重跑)</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/orchestrator/reflexion.py#L67">reflexion.py:67</a>
- </div>
- <div class="pill-row">
- <span class="tag llm">LLM 调用</span>
- <span class="tag degrade">置信度 < 0.7 触发重试</span>
- </div>
- <h4>触发条件</h4>
- <p>仅对 <code class="inline">Extractor</code> / <code class="inline">GapAnalyzer</code> / <code class="inline">InnovationGenerator</code> 启用。步骤完成后,让 LLM 对照上下文检查结果质量。</p>
- <details open>
- <summary>LLM Prompt · 自校验</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是质量审核员,负责对照上下文检查 Agent 输出质量。只输出 JSON。
- <span class="tok-user">[User]</span>
- 请对照上下文检查以下步骤的结果质量。
- 步骤:<span class="placeholder">{step_name}</span>
- 上下文:
- 步骤 2 结果:{抽取记录摘要...}
- 结果:
- <span class="placeholder">{result_str}</span>
- 检查要求:
- 1. 评估结果是否准确、完整、无臆造;
- 2. 输出 0-1 的置信度(confidence);
- 3. 列出发现的问题(issues);
- 4. 置信度低于 <span class="placeholder">{CONFIDENCE_THRESHOLD=0.7}</span> 时 needs_retry=true;
- 5. 指出需重试的字段或方面(retry_fields)。
- 严格以 JSON 输出:
- {"confidence": 0.85, "issues": ["问题1"], "needs_retry": false, "retry_fields": []}</pre>
- </details>
- <h4>重试逻辑</h4>
- <ol>
- <li><code class="inline">confidence < 0.7</code> → <code class="inline">needs_retry = true</code></li>
- <li>调用 <code class="inline">retry_func()</code> 重新执行该步骤(针对性补抽,非整篇重跑)</li>
- <li>最多重试 <code class="inline">MAX_RETRIES=2</code> 次</li>
- <li>仍不达标 → 标记 <code class="inline">needs_human_review</code>,流程继续不阻塞</li>
- </ol>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- <span class="tok-str">"step_name"</span>: <span class="tok-str">"GapAnalyzer"</span>,
- <span class="tok-str">"result"</span>: <span class="tok-com">{步骤结果}</span>,
- <span class="tok-str">"context"</span>: <span class="tok-com">{依赖步骤结果}</span>
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出</div>
- <pre>{
- "confidence": 0.85,
- "issues": ["..."],
- "needs_retry": false,
- "retry_fields": []
- }</pre>
- </div>
- </div>
- </div>
- </section>
- <!-- Step 5 -->
- <section id="step5">
- <div class="card step-card s5">
- <div class="step-header">
- <div>
- <span class="step-tag s5"><span class="step-num">⑤</span> 创新点生成</span>
- <span class="step-title">四类来源 + 三维评分筛选 + 联网去重</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/tools/generate_innovations.py#L58">generate_innovations.py:58</a>
- </div>
- <div class="pill-row">
- <span class="tag llm">LLM 调用</span>
- <span class="tag tool">web_search 工具(DuckDuckGo)</span>
- <span class="tag degrade">联网失败降级 skip_novelty_check</span>
- </div>
- <h4>处理流程</h4>
- <ol>
- <li><strong>调用 LLM</strong>:针对每个 Gap 从四类来源生成候选创新方向</li>
- <li><strong>规范化 + 排序</strong>:按三维评分之和降序</li>
- <li><strong>取 Top 2-3</strong></li>
- <li><strong>联网新颖性去重</strong>:对每个创新点用 DuckDuckGo 搜索近 1-2 年 arXiv/Google Scholar,验证是否已有相同工作</li>
- </ol>
- <details open>
- <summary>LLM Prompt · 创新生成</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是一位学术创新生成专家,擅长从研究空白中挖掘可执行的创新方向。
- 只输出严格 JSON,不要输出任何其他内容。
- <span class="tok-user">[User]</span>
- 你是一位富有创造力的学术研究者,需要基于以下研究空白(Gap)生成候选创新方向。
- ## 用户研究方向约束
- 用户的研究方向为:<span class="placeholder">{research_direction}</span>
- 生成的创新点应尽量与该方向相关,并在 supporting_evidence 中说明创新点与该方向的关联。
- ## 研究空白列表
- - [repeated_limitation] 位置编码外推性普遍较差(来源: Attention..., BERT...)
- - [method_gap] ...
- ## 任务
- 针对每个 Gap,从以下四类来源生成候选创新方向:
- 1. <span class="tok-kw">method_combination</span>(方法组合):把 A 论文方法迁移到 B 论文问题
- 2. <span class="tok-kw">limitation_fix</span>(局限改进):直接解决被反复提及的局限
- 3. <span class="tok-kw">cross_domain</span>(跨域迁移):引入其他领域方法
- 4. <span class="tok-kw">new_scenario</span>(新场景应用):把现有方法用到未被覆盖的数据/场景
- 对每个候选创新方向,用三维评分(0-10 分):
- - novelty(新颖性)
- - feasibility(可行性)
- - significance(显著性)
- ## 输出要求
- 严格输出 JSON:
- {
- "innovations": [
- {
- "title": "创新点标题",
- "idea": "创新思路详细描述",
- "source": "method_combination|limitation_fix|cross_domain|new_scenario",
- "gap_origin": "对应的 Gap 描述",
- "score": {"novelty": 8.0, "feasibility": 7.5, "significance": 8.0},
- "supporting_evidence": "论据支撑"
- }
- ]
- }
- 生成多个候选(每个 Gap 至少 1-2 个),按综合得分(三项之和)从高到低排序。</pre>
- </details>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- <span class="tok-str">"gaps"</span>: [<span class="tok-com">{Gap}</span>, ...],
- <span class="tok-str">"research_direction"</span>: <span class="tok-str">"多模态学习"</span>
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出</div>
- <pre>{
- "innovations": [
- {
- "title": "基于 RoPE 的多模态位置编码",
- "idea": "...",
- "source": "limitation_fix",
- "gap_origin": "位置编码外推性差",
- "score": {"novelty":8.0,"feasibility":7.5,"significance":8.5},
- "novelty_check": "近 1-2 年无相同工作",
- "supporting_evidence": "..."
- }
- ]
- }</pre>
- </div>
- </div>
- </div>
- </section>
- <!-- Step 6 -->
- <section id="step6">
- <div class="card step-card s6">
- <div class="step-header">
- <div>
- <span class="step-tag s6"><span class="step-num">⑥</span> 实验方案设计</span>
- <span class="step-title">8 要素可执行研究方案</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/tools/design_experiments.py#L34">design_experiments.py:34</a>
- </div>
- <div class="pill-row">
- <span class="tag llm">LLM 调用</span>
- </div>
- <details open>
- <summary>LLM Prompt · 实验设计</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是一位学术实验设计专家,擅长设计严谨可执行的研究方案。
- 只输出严格 JSON,不要输出任何其他内容。
- <span class="tok-user">[User]</span>
- 你是一位资深学术研究者,需要为以下创新点设计可执行的研究方案。
- ## 创新点列表
- - 基于 RoPE 的多模态位置编码: ...(来源: limitation_fix,Gap: 位置编码外推性差)
- ## 任务
- 为每个创新点设计完整、可执行的研究方案,包含:
- - <span class="tok-kw">hypothesis</span>(研究假设):清晰可验证的假设
- - <span class="tok-kw">datasets</span>(数据集列表):适合该研究的公开数据集
- - <span class="tok-kw">baselines</span>(基线方法列表):需要对比的现有方法
- - <span class="tok-kw">metrics</span>(评估指标列表):量化评估指标
- - <span class="tok-kw">ablation</span>(消融实验设计列表):验证各组件贡献
- - <span class="tok-kw">steps</span>(实验步骤列表):可复现的实验流程
- - <span class="tok-kw">expected_results</span>(预期结果):预期达到的性能或发现
- - <span class="tok-kw">risks</span>(风险):可能的失败点与应对
- ## 输出要求
- 严格输出 JSON:
- {
- "experiment_plans": [
- {
- "innovation_title": "对应的创新点标题",
- "hypothesis": "研究假设",
- "datasets": ["数据集1","数据集2"],
- "baselines": ["基线方法1","基线方法2"],
- "metrics": ["指标1","指标2"],
- "ablation": ["消融实验1"],
- "steps": ["步骤1","步骤2","步骤3"],
- "expected_results": "预期结果描述",
- "risks": "风险与应对描述"
- }
- ]
- }
- 每个创新点对应一条方案,方案须具体可执行。</pre>
- </details>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- <span class="tok-str">"innovations"</span>: [<span class="tok-com">{Innovation}</span>, ...]
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出</div>
- <pre>{
- "experiment_plans": [
- {
- "innovation_title": "基于 RoPE 的多模态位置编码",
- "hypothesis": "RoPE 能提升多模态任务外推性",
- "datasets": ["COCO","VQA"],
- "baselines": ["ALiBi"," vanilla PE"],
- "metrics": ["BLEU","Accuracy"],
- "ablation": ["移除旋转矩阵"],
- "steps": ["..."],
- "expected_results": "...",
- "risks": "..."
- }
- ]
- }</pre>
- </div>
- </div>
- </div>
- </section>
- <!-- Step 7 -->
- <section id="step7">
- <div class="card step-card s7">
- <div class="step-header">
- <div>
- <span class="step-tag s7"><span class="step-num">⑦</span> 整合输出</span>
- <span class="step-title">Markdown 研究提案 · 可追溯</span>
- </div>
- <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/agents/report_agent.py#L47">report_agent.py:47</a>
- </div>
- <div class="pill-row">
- <span class="tag llm">LLM 调用(max_tokens=8192)</span>
- <span class="tag degrade">LLM 失败降级本地格式化</span>
- </div>
- <h4>输入聚合 <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/orchestrator/workflow.py#L127">workflow.py:127</a></h4>
- <p><code class="inline">_build_final_results()</code> 从各步骤的 <code class="inline">tool_calls</code> 里聚合 4 字段:</p>
- <pre>{
- "papers_records": [...], <span class="tok-com"># 步骤②所有抽取记录</span>
- "gaps": [...], <span class="tok-com"># 步骤④所有 Gap</span>
- "innovations": [...], <span class="tok-com"># 步骤⑤所有创新点</span>
- "experiment_plans": [...] <span class="tok-com"># 步骤⑥所有实验方案</span>
- }</pre>
- <details open>
- <summary>LLM Prompt · 整合报告</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是学术研究提案撰写专家,擅长把多篇论文的 Gap 分析、创新点与实验方案
- 整合为结构清晰、可追溯的研究提案。严格输出 JSON,不要输出任何额外说明文本。
- <span class="tok-user">[User]</span>
- 请基于以下前序步骤产物,整合为一份结构化研究提案。
- ## 研究主题
- <span class="placeholder">{topic}</span>
- ## 一、论文结构化记录(含 source_sections,可追溯)
- ----- 论文 1: Attention Is All You Need (path=...) -----
- [研究任务/问题] 序列建模... (出处章节: Introduction)
- [核心方法] Transformer... (出处章节: 3 Method; 3.1)
- [局限性] 位置编码外推性差 (出处章节: 6 Conclusion)
- ...
- ## 二、识别到的 Gap 列表
- Gap 1 [repeated_limitation] (置信度=0.85): 位置编码外推性普遍较差
- 来源论文: Attention...; BERT...
- 论证: 两篇均提及...
- ## 三、创新点列表(含三维评分)
- 创新点 1: 基于 RoPE 的多模态位置编码 (来源=limitation_fix)
- 创新思路: ...
- 三维评分: 新颖性 8 / 可行性 7 / 显著性 9
- 新颖性验证: 近 1-2 年无相同工作
- 论据支撑: ...
- ## 四、实验方案列表
- 方案 1 (对应: 基于 RoPE 的多模态位置编码)
- 研究假设: ...
- 数据集: COCO, VQA
- ...
- ## 整合任务
- 1. 背景综述:基于 Gap 分析梳理领域现状与脉络
- 2. 创新点论证:对每个创新点补充 Gap 来源 + 创新思路 + 三维评分 + 论据支撑
- 3. 实验方案:保留每个创新点的完整实验方案要素
- 4. 参考文献溯源:所有引用论文的来源标注,可追溯至具体章节(source_sections)
- 5. Markdown 报告:按指定结构生成完整 Markdown 文本
- ## Markdown 报告结构要求
- # 研究提案:[主题]
- ## 一、背景综述
- ## 二、创新点论证
- ### 创新点 1:[标题]
- - Gap 来源:...
- - 创新思路:...
- - 三维评分:新颖性 X / 可行性 X / 显著性 X
- - 论据支撑:...
- ## 三、实验方案
- ## 四、参考文献溯源
- ## 输出要求
- 严格输出 JSON:
- {
- "background_review": "背景综述文本",
- "innovations": [...],
- "experiment_plans": [...],
- "references": [{"title":"...","path":"...","sections":[...]}],
- "markdown": "完整 Markdown 报告字符串"
- }
- 注意:
- - references 必须可追溯至具体章节(source_sections)
- - 不得编造论文或证据,所有引用须来自上述论文记录
- - 若某字段在输入中缺失,对应位置写"暂无"而非编造</pre>
- </details>
- <div class="io-grid">
- <div class="io-box in">
- <div class="label">输入</div>
- <pre>{
- "papers_records": [...],
- "gaps": [...],
- "innovations": [...],
- "experiment_plans": [...]
- }</pre>
- </div>
- <div class="io-box out">
- <div class="label">输出</div>
- <pre>{
- "background_review": "...",
- "innovations": [...],
- "experiment_plans": [...],
- "references": [
- {"title":"...","path":"...",
- "sections":["3 Method","6 Conclusion"]}
- ],
- "markdown": "# 研究提案:...\n## 一、背景综述\n..."
- }</pre>
- </div>
- </div>
- </div>
- </section>
- <!-- Evaluation -->
- <section id="eval">
- <h2 class="section-title"><span class="num">6</span>评估阶段</h2>
- <p class="section-desc">流水线完成后,<code class="inline">Evaluator</code> 汇总工具调用追踪 + 输出质量评估,产出统一评估结果。</p>
- <div class="card">
- <h4>调用位置 <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/evaluator/evaluator.py#L28">evaluator.py:28</a></h4>
- <ol>
- <li><code class="inline">ToolCallTracker</code> 汇总工具调用成功率 / 总数 / 失败数 / 耗时</li>
- <li><code class="inline">QualityEvaluator</code> 对每个创新点调 LLM 做三维评分校验</li>
- </ol>
- <details open>
- <summary>LLM Prompt · 质量评估</summary>
- <pre class="prompt"><span class="tok-sys">[System]</span>
- 你是一位学术评审专家,擅长对创新点进行多维评分。只输出严格 JSON。
- <span class="tok-user">[User]</span>
- 你是一位严谨的学术评审专家,请对以下创新点进行三维评分(0-10 分,保留一位小数)。
- ## 创新点信息
- - 标题:<span class="placeholder">{title}</span>
- - 思路:<span class="placeholder">{idea}</span>
- - 论据支撑:<span class="placeholder">{supporting_evidence}</span>
- - 新颖性检查:<span class="placeholder">{novelty_check}</span>
- ## 评分维度
- - <span class="tok-kw">novelty</span>(新颖性):与现有工作的差异程度
- - <span class="tok-kw">feasibility</span>(可行性):技术实现与资源获取的难度
- - <span class="tok-kw">significance</span>(显著性):学术或应用价值
- ## 输出要求
- 严格输出 JSON:
- {
- "innovation_title": "...",
- "score": {"novelty": 8.0, "feasibility": 7.5, "significance": 8.0},
- "reasoning": "评分理由,简明扼要"
- }</pre>
- </details>
- <h4>输出</h4>
- <pre>{
- "innovations_scores": [...],
- "tool_call_success_rate": 0.89,
- "total_tool_calls": 47,
- "failed_tool_calls": 5,
- "avg_confidence": 0.78,
- "tool_calls": [...],
- "quality_assessment": {...}
- }</pre>
- </div>
- </section>
- <!-- Summary -->
- <section id="summary">
- <h2 class="section-title"><span class="num">7</span>完整 LLM 调用清单</h2>
- <p class="section-desc">单次完整运行共触发 11 次 LLM 调用(不含工具内部重试)。</p>
- <table>
- <thead>
- <tr>
- <th>#</th>
- <th>阶段</th>
- <th>调用位置</th>
- <th>Prompt 设计要点</th>
- <th>输出格式</th>
- </tr>
- </thead>
- <tbody>
- <tr><td>1</td><td>Plan</td><td>controller.py:151</td><td>Agent 编排器,生成动态计划</td><td>JSON 数组</td></tr>
- <tr><td>2</td><td>步骤②</td><td>extract_schema.py:102</td><td>强制 Function Calling,source_sections 防编造,limitations 严禁臆造</td><td>FieldSource JSON</td></tr>
- <tr><td>3</td><td>步骤②补抽</td><td>extract_paper_structure.py:161</td><td>只补缺失字段,不重跑整篇</td><td>FieldSource JSON</td></tr>
- <tr><td>4</td><td>步骤②自校验</td><td>extract_paper_structure.py:306</td><td>对照原文逐字段检查置信度</td><td>JSON</td></tr>
- <tr><td>5</td><td>步骤④</td><td>analyze_gaps.py:113</td><td>四类 Gap 信号识别,对比矩阵输入</td><td>JSON gaps 列表</td></tr>
- <tr><td>6</td><td>步骤④ Reflexion</td><td>reflexion.py:67</td><td>质量审核员,对照上下文检查</td><td>JSON confidence</td></tr>
- <tr><td>7</td><td>步骤⑤</td><td>generate_innovations.py:81</td><td>四类创新方向 + 三维评分 + 方向约束</td><td>JSON innovations</td></tr>
- <tr><td>8</td><td>步骤⑤ Reflexion</td><td>reflexion.py:67</td><td>同上</td><td>JSON confidence</td></tr>
- <tr><td>9</td><td>步骤⑥</td><td>design_experiments.py:51</td><td>8 要素实验方案</td><td>JSON plans</td></tr>
- <tr><td>10</td><td>步骤⑦</td><td>report_agent.py:197</td><td>整合 4 字段为 Markdown 研究提案</td><td>JSON + markdown</td></tr>
- <tr><td>11</td><td>评估</td><td>quality.py:73</td><td>学术评审专家三维评分校验</td><td>JSON score</td></tr>
- </tbody>
- </table>
- <h3>工具调用清单</h3>
- <table>
- <thead>
- <tr><th>工具</th><th>步骤</th><th>输入</th><th>输出</th><th>特性</th></tr>
- </thead>
- <tbody>
- <tr><td><code class="inline">scan_folder</code></td><td>①</td><td>folder_path</td><td>pdf_files 列表</td><td>本地</td></tr>
- <tr><td><code class="inline">parse_pdf</code></td><td>①</td><td>file_path</td><td>ParsedPaper</td><td>本地 · 并发 · 缓存</td></tr>
- <tr><td><code class="inline">extract_paper_structure</code></td><td>②</td><td>paper_path, parsed_paper</td><td>PaperRecord</td><td>LLM · 并发 · 缓存</td></tr>
- <tr><td><code class="inline">build_vector_index</code></td><td>③</td><td>papers_json</td><td>indexed_count</td><td>本地 · 持久化</td></tr>
- <tr><td><code class="inline">analyze_gaps</code></td><td>④</td><td>papers_records</td><td>gaps 列表</td><td>LLM · 内部调 RAG</td></tr>
- <tr><td><code class="inline">hybrid_search</code></td><td>④⑤</td><td>query</td><td>检索片段 + 分数</td><td>Chroma + BM25</td></tr>
- <tr><td><code class="inline">generate_innovations</code></td><td>⑤</td><td>gaps, research_direction</td><td>innovations 列表</td><td>LLM · 内部调 web_search</td></tr>
- <tr><td><code class="inline">web_search</code></td><td>⑤</td><td>query</td><td>搜索结果摘要</td><td>DuckDuckGo</td></tr>
- <tr><td><code class="inline">design_experiments</code></td><td>⑥</td><td>innovations</td><td>experiment_plans</td><td>LLM</td></tr>
- <tr><td><code class="inline">load_memory / save_memory</code></td><td>⑤</td><td>type, query</td><td>历史 Gap/创新点</td><td>Chroma 长期记忆</td></tr>
- </tbody>
- </table>
- </section>
- <!-- Principles -->
- <section id="principles">
- <h2 class="section-title"><span class="num">8</span>Prompt 设计的 5 个核心原则</h2>
- <p class="section-desc">贯穿所有 prompt 的设计哲学。</p>
- <div class="card">
- <h4>① System 角色锚定</h4>
- <p>每个 prompt 都用 system 消息锁定角色("你是 XX 专家"),并在末尾强调"只输出严格 JSON",避免自由文本。</p>
- <h4>② 结构化输出强制</h4>
- <p>全部用 Function Calling 或显式 JSON Schema,避免 LLM 自由发挥导致解析失败。步骤② 用 <code class="inline">tool_choice={"type":"function",...}</code> 强制调用。</p>
- <h4>③ 防编造机制</h4>
- <p><code class="inline">source_sections</code> 强制溯源 + "严禁臆造" + "缺失填暂无"。步骤② limitations 字段特别约束:论文未明确承认时填"论文未明确陈述局限"而非编造。</p>
- <h4>④ 上下文注入</h4>
- <p>把前序步骤的结构化产物作为"证据"喂给模型,而非让模型凭空推理。步骤④ 喂对比矩阵、步骤⑦ 喂 4 字段聚合。</p>
- <h4>⑤ 降级兜底</h4>
- <p>每个 LLM 调用都有 <code class="inline">try/except</code>,失败时回退到本地格式化或默认值。步骤⑦ LLM 失败 → <code class="inline">_format_markdown()</code> 本地组装。</p>
- </div>
- </section>
- <!-- Degradation -->
- <section id="degrade">
- <h2 class="section-title"><span class="num">9</span>降级与容错策略</h2>
- <p class="section-desc">每个步骤都有失败兜底,局部失败不阻塞全局。</p>
- <table>
- <thead>
- <tr><th>失败点</th><th>降级路径</th><th>影响</th></tr>
- </thead>
- <tbody>
- <tr><td>单篇论文解析失败</td><td>隔离跳过,独立缓存可单独重跑</td><td>不影响其他论文</td></tr>
- <tr><td>LLM 输出非 JSON</td><td>Function Calling 强制 schema + <code class="inline">parse_json_safe</code> 程序层兜底</td><td>该字段降级为空</td></tr>
- <tr><td>工具调用超时/失败</td><td>3 次重试 + 降级路径</td><td>记录 <code class="inline">degraded_steps</code></td></tr>
- <tr><td>论文超长截断</td><td>按章节分块抽取后合并去重</td><td>无影响</td></tr>
- <tr><td>Controller 规划失败</td><td>异常捕获 + 回退默认七步</td><td>无影响</td></tr>
- <tr><td>向量索引构建失败</td><td>跳过 RAG,仅用结构化字段做 Gap 分析</td><td>Gap evidence 无原文佐证</td></tr>
- <tr><td>联网搜索失败</td><td>跳过新颖性去重,标记"未验证"</td><td>novelty_check 为空</td></tr>
- <tr><td>Reflexion 重试 2 次仍不达标</td><td>标记 <code class="inline">needs_human_review</code>,流程继续</td><td>评估报告标注</td></tr>
- <tr><td>步骤⑦ LLM 整合失败</td><td>本地 <code class="inline">_format_markdown()</code> 格式化组装</td><td>报告样式简化</td></tr>
- <tr><td>LLM 编造内容</td><td>所有字段强制 <code class="inline">source_sections</code> 字段级回溯</td><td>可人工核查</td></tr>
- </tbody>
- </table>
- <div class="callout ok">
- <strong>设计哲学:</strong>"工具即思维" — Function Calling 把"决策"显式化为工具调用,可观测、可回放;"降级即韧性" — 每个步骤都有失败兜底,局部失败不阻塞全局。
- </div>
- </section>
- <div class="footer">
- PaperMind · 论文创新点发现 Agent · 端到端运行流程详解<br>
- 基于源码 <code style="background:rgba(255,255,255,0.05);padding:2px 6px;border-radius:4px;">backend/</code> 自动生成 · 共 11 次 LLM 调用 · 10 个工具
- </div>
- </main>
- </div>
- <script>
- // 侧边栏高亮当前 section
- const links = document.querySelectorAll('.sidebar a');
- const sections = document.querySelectorAll('section[id], div[id]');
- const observer = new IntersectionObserver((entries) => {
- entries.forEach(entry => {
- if (entry.isIntersecting) {
- const id = entry.target.id;
- links.forEach(l => {
- l.classList.toggle('active', l.getAttribute('href') === '#' + id);
- });
- }
- });
- }, { rootMargin: '-20% 0px -70% 0px' });
- sections.forEach(s => observer.observe(s));
- </script>
- </body>
- </html>
|