pipeline_walkthrough.html 64 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426
  1. <!DOCTYPE html>
  2. <html lang="zh-CN">
  3. <head>
  4. <meta charset="UTF-8">
  5. <meta name="viewport" content="width=device-width, initial-scale=1.0">
  6. <title>PaperMind · 端到端运行流程详解</title>
  7. <style>
  8. :root {
  9. --bg: #0f1117;
  10. --bg-elev: #161924;
  11. --bg-card: #1c2030;
  12. --bg-code: #0a0c14;
  13. --border: #2a2f42;
  14. --text: #e6e9f0;
  15. --text-dim: #9aa0b0;
  16. --text-mute: #6b7180;
  17. --accent: #5b9dff;
  18. --accent-2: #a78bfa;
  19. --green: #4ade80;
  20. --yellow: #fbbf24;
  21. --red: #f87171;
  22. --cyan: #22d3ee;
  23. --pink: #f472b6;
  24. --orange: #fb923c;
  25. --step1: #60a5fa;
  26. --step2: #a78bfa;
  27. --step3: #22d3ee;
  28. --step4: #4ade80;
  29. --step5: #fbbf24;
  30. --step6: #fb923c;
  31. --step7: #f472b6;
  32. }
  33. * { box-sizing: border-box; margin: 0; padding: 0; }
  34. html { scroll-behavior: smooth; }
  35. body {
  36. font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC", "Microsoft YaHei", sans-serif;
  37. background: var(--bg);
  38. color: var(--text);
  39. line-height: 1.65;
  40. font-size: 15px;
  41. }
  42. /* ===== Layout ===== */
  43. .layout { display: flex; max-width: 1480px; margin: 0 auto; }
  44. .sidebar {
  45. position: sticky; top: 0; height: 100vh; width: 260px; flex-shrink: 0;
  46. border-right: 1px solid var(--border);
  47. background: var(--bg-elev);
  48. overflow-y: auto;
  49. padding: 24px 16px;
  50. }
  51. .sidebar h2 { font-size: 13px; color: var(--text-mute); text-transform: uppercase; letter-spacing: 1px; margin-bottom: 12px; padding: 0 8px; }
  52. .sidebar a {
  53. display: block; padding: 7px 12px; color: var(--text-dim); text-decoration: none;
  54. border-radius: 6px; font-size: 13.5px; transition: all 0.15s;
  55. border-left: 2px solid transparent;
  56. }
  57. .sidebar a:hover { color: var(--text); background: rgba(91,157,255,0.08); }
  58. .sidebar a.active { color: var(--accent); background: rgba(91,157,255,0.12); border-left-color: var(--accent); }
  59. .sidebar a.sub { padding-left: 24px; font-size: 12.5px; }
  60. .main { flex: 1; min-width: 0; padding: 48px 56px 120px; }
  61. /* ===== Hero ===== */
  62. .hero {
  63. background: linear-gradient(135deg, #1a1f3a 0%, #0f1117 60%);
  64. border: 1px solid var(--border);
  65. border-radius: 16px;
  66. padding: 48px 44px;
  67. margin-bottom: 48px;
  68. position: relative;
  69. overflow: hidden;
  70. }
  71. .hero::before {
  72. content: ""; position: absolute; top: -50%; right: -10%; width: 500px; height: 500px;
  73. background: radial-gradient(circle, rgba(91,157,255,0.15), transparent 70%);
  74. pointer-events: none;
  75. }
  76. .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; }
  77. .hero .sub { font-size: 16px; color: var(--text-dim); max-width: 760px; }
  78. .hero .meta { margin-top: 24px; display: flex; gap: 12px; flex-wrap: wrap; }
  79. .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; }
  80. /* ===== Section ===== */
  81. section { margin-bottom: 56px; scroll-margin-top: 20px; }
  82. h2.section-title {
  83. font-size: 24px; font-weight: 700; margin-bottom: 8px;
  84. padding-bottom: 12px; border-bottom: 1px solid var(--border);
  85. display: flex; align-items: center; gap: 10px;
  86. }
  87. h2.section-title .num {
  88. display: inline-flex; align-items: center; justify-content: center;
  89. width: 32px; height: 32px; border-radius: 8px;
  90. background: linear-gradient(135deg, var(--accent), var(--accent-2));
  91. color: #fff; font-size: 15px; font-weight: 700;
  92. }
  93. .section-desc { color: var(--text-dim); margin-bottom: 24px; font-size: 14.5px; }
  94. h3 { font-size: 18px; font-weight: 600; margin: 28px 0 14px; color: var(--text); }
  95. h4 { font-size: 15px; font-weight: 600; margin: 18px 0 10px; color: var(--accent); }
  96. p { margin-bottom: 10px; }
  97. /* ===== Cards ===== */
  98. .card {
  99. background: var(--bg-card);
  100. border: 1px solid var(--border);
  101. border-radius: 12px;
  102. padding: 22px 24px;
  103. margin-bottom: 18px;
  104. }
  105. .card.glow { box-shadow: 0 0 0 1px rgba(91,157,255,0.15), 0 8px 32px rgba(0,0,0,0.3); }
  106. .step-card { border-left: 3px solid var(--accent); }
  107. .step-card.s1 { border-left-color: var(--step1); }
  108. .step-card.s2 { border-left-color: var(--step2); }
  109. .step-card.s3 { border-left-color: var(--step3); }
  110. .step-card.s4 { border-left-color: var(--step4); }
  111. .step-card.s5 { border-left-color: var(--step5); }
  112. .step-card.s6 { border-left-color: var(--step6); }
  113. .step-card.s7 { border-left-color: var(--step7); }
  114. .step-header { display: flex; align-items: center; justify-content: space-between; margin-bottom: 14px; flex-wrap: wrap; gap: 10px; }
  115. .step-tag {
  116. display: inline-flex; align-items: center; gap: 8px;
  117. padding: 5px 12px; border-radius: 6px; font-size: 13px; font-weight: 600;
  118. }
  119. .step-tag.s1 { background: rgba(96,165,250,0.15); color: var(--step1); }
  120. .step-tag.s2 { background: rgba(167,139,250,0.15); color: var(--step2); }
  121. .step-tag.s3 { background: rgba(34,211,238,0.15); color: var(--step3); }
  122. .step-tag.s4 { background: rgba(74,222,128,0.15); color: var(--step4); }
  123. .step-tag.s5 { background: rgba(251,191,36,0.15); color: var(--step5); }
  124. .step-tag.s6 { background: rgba(251,146,60,0.15); color: var(--step6); }
  125. .step-tag.s7 { background: rgba(244,114,182,0.15); color: var(--step7); }
  126. .step-num { font-size: 18px; font-weight: 700; }
  127. .step-title { font-size: 17px; font-weight: 600; }
  128. .code-link {
  129. font-family: "JetBrains Mono", Consolas, monospace; font-size: 12px;
  130. color: var(--text-dim); text-decoration: none;
  131. background: rgba(0,0,0,0.3); padding: 3px 8px; border-radius: 4px;
  132. border: 1px solid var(--border);
  133. }
  134. .code-link:hover { color: var(--accent); border-color: var(--accent); }
  135. /* ===== IO blocks ===== */
  136. .io-grid { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; margin: 14px 0; }
  137. @media (max-width: 900px) { .io-grid { grid-template-columns: 1fr; } }
  138. .io-box {
  139. background: var(--bg-code); border: 1px solid var(--border); border-radius: 8px;
  140. padding: 14px 16px;
  141. }
  142. .io-box .label {
  143. font-size: 11px; font-weight: 700; text-transform: uppercase; letter-spacing: 1px;
  144. margin-bottom: 8px; display: flex; align-items: center; gap: 6px;
  145. }
  146. .io-box.in .label { color: var(--cyan); }
  147. .io-box.out .label { color: var(--green); }
  148. .io-box .label::before { content: ""; width: 8px; height: 8px; border-radius: 50%; }
  149. .io-box.in .label::before { background: var(--cyan); }
  150. .io-box.out .label::before { background: var(--green); }
  151. /* ===== Code blocks ===== */
  152. pre {
  153. background: var(--bg-code); border: 1px solid var(--border); border-radius: 8px;
  154. padding: 14px 16px; overflow-x: auto; font-size: 12.5px; line-height: 1.55;
  155. font-family: "JetBrains Mono", Consolas, "Courier New", monospace;
  156. color: #d4d4dc;
  157. }
  158. pre.prompt {
  159. background: linear-gradient(180deg, #14111f 0%, #0a0c14 100%);
  160. border-color: rgba(167,139,250,0.3);
  161. }
  162. code.inline {
  163. background: rgba(91,157,255,0.12); color: var(--accent);
  164. padding: 1px 6px; border-radius: 4px; font-size: 12.5px;
  165. font-family: "JetBrains Mono", Consolas, monospace;
  166. }
  167. .tok-sys { color: #c792ea; }
  168. .tok-user { color: #82aaff; }
  169. .tok-kw { color: #c792ea; }
  170. .tok-str { color: #c3e88d; }
  171. .tok-com { color: #6b7180; font-style: italic; }
  172. .tok-num { color: #f78c6c; }
  173. /* ===== Collapsible ===== */
  174. details {
  175. background: var(--bg-code); border: 1px solid var(--border); border-radius: 8px;
  176. margin: 10px 0; overflow: hidden;
  177. }
  178. details[open] { border-color: rgba(167,139,250,0.4); }
  179. summary {
  180. cursor: pointer; padding: 12px 16px; font-size: 13.5px; font-weight: 600;
  181. color: var(--accent-2); user-select: none;
  182. display: flex; align-items: center; gap: 8px;
  183. background: linear-gradient(90deg, rgba(167,139,250,0.08), transparent);
  184. }
  185. summary:hover { background: rgba(167,139,250,0.1); }
  186. summary::before { content: "▸"; transition: transform 0.15s; font-size: 11px; }
  187. details[open] summary::before { transform: rotate(90deg); }
  188. details > pre, details > div { border: none; border-radius: 0; margin: 0; border-top: 1px solid var(--border); }
  189. /* ===== Tables ===== */
  190. table {
  191. width: 100%; border-collapse: collapse; margin: 14px 0;
  192. background: var(--bg-card); border: 1px solid var(--border); border-radius: 8px;
  193. overflow: hidden; font-size: 13px;
  194. }
  195. th, td { padding: 10px 14px; text-align: left; border-bottom: 1px solid var(--border); vertical-align: top; }
  196. th { background: rgba(91,157,255,0.06); color: var(--accent); font-weight: 600; font-size: 12px; text-transform: uppercase; letter-spacing: 0.5px; }
  197. tr:last-child td { border-bottom: none; }
  198. tr:hover td { background: rgba(91,157,255,0.04); }
  199. /* ===== Pipeline diagram ===== */
  200. .pipeline {
  201. display: flex; flex-direction: column; gap: 0; margin: 24px 0;
  202. }
  203. .pipeline-row {
  204. display: flex; align-items: stretch; gap: 0;
  205. }
  206. .pipe-node {
  207. flex: 1; background: var(--bg-card); border: 1px solid var(--border);
  208. border-radius: 10px; padding: 14px; position: relative;
  209. transition: transform 0.15s;
  210. }
  211. .pipe-node:hover { transform: translateY(-2px); }
  212. .pipe-node .pn-num {
  213. font-size: 11px; font-weight: 700; color: var(--text-mute); margin-bottom: 4px;
  214. }
  215. .pipe-node .pn-title { font-size: 14px; font-weight: 600; margin-bottom: 4px; }
  216. .pipe-node .pn-desc { font-size: 11.5px; color: var(--text-dim); line-height: 1.5; }
  217. .pipe-node .pn-llm { font-size: 10.5px; margin-top: 6px; color: var(--accent-2); }
  218. .pipe-arrow {
  219. display: flex; align-items: center; justify-content: center;
  220. width: 28px; color: var(--text-mute); font-size: 18px;
  221. }
  222. /* ===== Tags ===== */
  223. .tag {
  224. display: inline-block; padding: 2px 8px; border-radius: 4px;
  225. font-size: 11px; font-weight: 600; margin-right: 4px;
  226. }
  227. .tag.llm { background: rgba(167,139,250,0.15); color: var(--accent-2); }
  228. .tag.tool { background: rgba(34,211,238,0.15); color: var(--cyan); }
  229. .tag.local { background: rgba(74,222,128,0.15); color: var(--green); }
  230. .tag.cache { background: rgba(251,191,36,0.15); color: var(--yellow); }
  231. .tag.degrade { background: rgba(248,113,113,0.15); color: var(--red); }
  232. .tag.parallel { background: rgba(244,114,182,0.15); color: var(--pink); }
  233. /* ===== List ===== */
  234. ul, ol { margin: 8px 0 12px 22px; }
  235. li { margin-bottom: 5px; }
  236. li code.inline { font-size: 12px; }
  237. .callout {
  238. background: rgba(91,157,255,0.06); border-left: 3px solid var(--accent);
  239. padding: 12px 16px; border-radius: 0 8px 8px 0; margin: 12px 0; font-size: 13.5px;
  240. }
  241. .callout.warn { background: rgba(251,191,36,0.06); border-color: var(--yellow); }
  242. .callout.danger { background: rgba(248,113,113,0.06); border-color: var(--red); }
  243. .callout.ok { background: rgba(74,222,128,0.06); border-color: var(--green); }
  244. .flow-arrow { color: var(--text-mute); text-align: center; font-size: 18px; margin: 6px 0; }
  245. .pill-row { display: flex; flex-wrap: wrap; gap: 6px; margin: 8px 0; }
  246. .keyword { color: #c3e88d; }
  247. .placeholder { color: #f78c6c; }
  248. .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; }
  249. ::-webkit-scrollbar { width: 10px; height: 10px; }
  250. ::-webkit-scrollbar-track { background: var(--bg); }
  251. ::-webkit-scrollbar-thumb { background: var(--border); border-radius: 5px; }
  252. ::-webkit-scrollbar-thumb:hover { background: #3a4060; }
  253. </style>
  254. </head>
  255. <body>
  256. <div class="layout">
  257. <!-- Sidebar -->
  258. <nav class="sidebar">
  259. <h2>导航</h2>
  260. <a href="#overview">总览</a>
  261. <a href="#startup">1. 启动阶段</a>
  262. <a href="#submit">2. 提交任务</a>
  263. <a href="#phase1">3. Phase 1 预处理</a>
  264. <a href="#step1" class="sub">① PDF 解析</a>
  265. <a href="#step2" class="sub">② 结构化抽取</a>
  266. <a href="#step3" class="sub">③ 向量索引</a>
  267. <a href="#phase2">4. Phase 2 Plan</a>
  268. <a href="#phase3">5. Phase 3 Execute</a>
  269. <a href="#step4" class="sub">④ Gap 识别</a>
  270. <a href="#reflexion" class="sub">↳ Reflexion</a>
  271. <a href="#step5" class="sub">⑤ 创新点生成</a>
  272. <a href="#step6" class="sub">⑥ 实验设计</a>
  273. <a href="#step7" class="sub">⑦ 整合输出</a>
  274. <a href="#eval">6. 评估阶段</a>
  275. <a href="#summary">7. LLM 调用清单</a>
  276. <a href="#principles">8. Prompt 设计原则</a>
  277. <a href="#degrade">9. 降级与容错</a>
  278. </nav>
  279. <!-- Main -->
  280. <main class="main">
  281. <!-- Hero -->
  282. <div class="hero">
  283. <h1>PaperMind · 端到端运行流程详解</h1>
  284. <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>
  285. <div class="meta">
  286. <span class="badge">Plan-and-Execute + Reflexion</span>
  287. <span class="badge">7 步漏斗式信息收敛</span>
  288. <span class="badge">DeepSeek Function Calling</span>
  289. <span class="badge">BGE-m3 + Chroma 混合检索</span>
  290. </div>
  291. </div>
  292. <!-- Overview -->
  293. <section id="overview">
  294. <h2 class="section-title"><span class="num">0</span>总览:七步漏斗流水线</h2>
  295. <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>
  296. <div class="pipeline">
  297. <div class="pipeline-row">
  298. <div class="pipe-node" style="border-left:3px solid var(--step1)">
  299. <div class="pn-num">STEP 1</div>
  300. <div class="pn-title">PDF 解析</div>
  301. <div class="pn-desc">PyMuPDF 提取章节结构</div>
  302. <div class="pn-llm"><span class="tag local">本地</span></div>
  303. </div>
  304. <div class="pipe-arrow">→</div>
  305. <div class="pipe-node" style="border-left:3px solid var(--step2)">
  306. <div class="pn-num">STEP 2</div>
  307. <div class="pn-title">结构化抽取</div>
  308. <div class="pn-desc">Function Calling 抽 9 字段</div>
  309. <div class="pn-llm"><span class="tag llm">LLM</span> <span class="tag parallel">并行</span></div>
  310. </div>
  311. <div class="pipe-arrow">→</div>
  312. <div class="pipe-node" style="border-left:3px solid var(--step3)">
  313. <div class="pn-num">STEP 3</div>
  314. <div class="pn-title">向量索引</div>
  315. <div class="pn-desc">BGE-m3 → Chroma + BM25</div>
  316. <div class="pn-llm"><span class="tag local">本地</span></div>
  317. </div>
  318. </div>
  319. <div class="flow-arrow">↓</div>
  320. <div class="pipeline-row">
  321. <div class="pipe-node" style="border-left:3px solid var(--step4)">
  322. <div class="pn-num">STEP 4</div>
  323. <div class="pn-title">Gap 识别</div>
  324. <div class="pn-desc">四类 Gap 信号综合推理</div>
  325. <div class="pn-llm"><span class="tag llm">LLM</span> <span class="tag tool">RAG</span></div>
  326. </div>
  327. <div class="pipe-arrow">→</div>
  328. <div class="pipe-node" style="border-left:3px solid var(--step5)">
  329. <div class="pn-num">STEP 5</div>
  330. <div class="pn-title">创新点生成</div>
  331. <div class="pn-desc">四类来源 + 三维评分筛选</div>
  332. <div class="pn-llm"><span class="tag llm">LLM</span> <span class="tag tool">联网</span></div>
  333. </div>
  334. <div class="pipe-arrow">→</div>
  335. <div class="pipe-node" style="border-left:3px solid var(--step6)">
  336. <div class="pn-num">STEP 6</div>
  337. <div class="pn-title">实验设计</div>
  338. <div class="pn-desc">8 要素可执行方案</div>
  339. <div class="pn-llm"><span class="tag llm">LLM</span></div>
  340. </div>
  341. <div class="pipe-arrow">→</div>
  342. <div class="pipe-node" style="border-left:3px solid var(--step7)">
  343. <div class="pn-num">STEP 7</div>
  344. <div class="pn-title">整合输出</div>
  345. <div class="pn-desc">Markdown 研究提案</div>
  346. <div class="pn-llm"><span class="tag llm">LLM</span></div>
  347. </div>
  348. </div>
  349. </div>
  350. <div class="callout">
  351. <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。
  352. </div>
  353. </section>
  354. <!-- Startup -->
  355. <section id="startup">
  356. <h2 class="section-title"><span class="num">1</span>启动阶段(进程级)</h2>
  357. <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>
  358. <div class="card">
  359. <h4>启动命令</h4>
  360. <pre>python run_backend.py
  361. # 等价于:uvicorn backend.api.main:app --host 0.0.0.0 --port 8000</pre>
  362. <h4>启动钩子(on_startup)</h4>
  363. <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>
  364. <pre><span class="tok-com"># 顺序预加载</span>
  365. import torch <span class="tok-com"># ~5s</span>
  366. from transformers import AutoModel, AutoTokenizer <span class="tok-com"># ~3s</span>
  367. import chromadb <span class="tok-com"># ~2s</span>
  368. <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>
  369. _embedder = get_default_embedder()
  370. _embedder.embed_texts([<span class="tok-str">"预加载测试文本"</span>])</pre>
  371. <div class="io-grid">
  372. <div class="io-box in">
  373. <div class="label">输入</div>
  374. <div>无(进程启动)</div>
  375. </div>
  376. <div class="io-box out">
  377. <div class="label">输出</div>
  378. <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>
  379. </div>
  380. </div>
  381. </div>
  382. </section>
  383. <!-- Submit -->
  384. <section id="submit">
  385. <h2 class="section-title"><span class="num">2</span>用户提交分析任务</h2>
  386. <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>
  387. <div class="card">
  388. <h4>HTTP 请求</h4>
  389. <pre><span class="tok-com"># POST /api/analyze</span>
  390. {
  391. <span class="tok-str">"folder_path"</span>: <span class="tok-str">"D:/papers"</span>,
  392. <span class="tok-str">"research_direction"</span>: <span class="tok-str">"多模态学习"</span>
  393. }</pre>
  394. <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>
  395. <ol>
  396. <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>
  397. <li><code class="inline">asyncio.create_task(run_analysis(...))</code> → 后台异步启动,<strong>不阻塞响应</strong></li>
  398. <li>立即返回 <code class="inline">{"task_id": "..."}</code></li>
  399. </ol>
  400. <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>
  401. <pre><span class="tok-com"># 初始化组件</span>
  402. client = DeepSeekClient() <span class="tok-com"># LLM 客户端</span>
  403. registry = build_default_registry() <span class="tok-com"># 工具注册表</span>
  404. memory = MemoryManager() <span class="tok-com"># 双层记忆</span>
  405. orchestrator = WorkflowOrchestrator(client, registry, memory)
  406. <span class="tok-com"># 执行流水线</span>
  407. pipeline_output = await orchestrator.run_pipeline(...)
  408. <span class="tok-com"># 生成报告 + 评估</span>
  409. report = await ReportAgent(client).generate_report(final_results)
  410. evaluation = await Evaluator(client).evaluate(...)
  411. <span class="tok-com"># 存结果</span>
  412. task_manager.update_task(task_id, status=<span class="tok-str">"completed"</span>, result={...})</pre>
  413. <div class="callout ok">
  414. <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。
  415. </div>
  416. </div>
  417. </section>
  418. <!-- Phase 1 -->
  419. <section id="phase1">
  420. <h2 class="section-title"><span class="num">3</span>Phase 1:预处理(步骤 1-3,带缓存)</h2>
  421. <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>
  422. </section>
  423. <!-- Step 1 -->
  424. <section id="step1">
  425. <div class="card step-card s1">
  426. <div class="step-header">
  427. <div>
  428. <span class="step-tag s1"><span class="step-num">①</span> PDF 解析</span>
  429. <span class="step-title">PyMuPDF 提取章节结构</span>
  430. </div>
  431. <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/tools/parse_pdf.py">parse_pdf.py</a>
  432. </div>
  433. <div class="pill-row">
  434. <span class="tag local">本地处理</span>
  435. <span class="tag parallel">并发 asyncio.gather</span>
  436. <span class="tag cache">缓存 data/cache/parsed/{paper_id}.json</span>
  437. </div>
  438. <h4>处理流程</h4>
  439. <ul>
  440. <li>PyMuPDF 提取每页文本 + 字号/字体信息</li>
  441. <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>
  442. <li>去页眉页脚(短行 &lt; 80 字符且重复出现)</li>
  443. <li>提取图表标题(<code class="inline">Figure X</code> / <code class="inline">表 X</code> 正则)</li>
  444. <li>合并跨页段落,修复断句</li>
  445. </ul>
  446. <div class="io-grid">
  447. <div class="io-box in">
  448. <div class="label">输入</div>
  449. <pre>{
  450. <span class="tok-str">"file_path"</span>: <span class="tok-str">"D:/papers/xxx.pdf"</span>
  451. }</pre>
  452. </div>
  453. <div class="io-box out">
  454. <div class="label">输出:ParsedPaper</div>
  455. <pre>{
  456. <span class="tok-str">"path"</span>: <span class="tok-str">"D:/papers/xxx.pdf"</span>,
  457. <span class="tok-str">"title"</span>: <span class="tok-str">"Attention Is All You Need"</span>,
  458. <span class="tok-str">"sections"</span>: [
  459. {
  460. <span class="tok-str">"heading"</span>: <span class="tok-str">"Abstract"</span>,
  461. <span class="tok-str">"content"</span>: <span class="tok-str">"..."</span>,
  462. <span class="tok-str">"figure_captions"</span>: [<span class="tok-str">"Figure 1: ..."</span>]
  463. }
  464. ]
  465. }</pre>
  466. </div>
  467. </div>
  468. <div class="callout">
  469. <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。
  470. </div>
  471. </div>
  472. </section>
  473. <!-- Step 2 -->
  474. <section id="step2">
  475. <div class="card step-card s2">
  476. <div class="step-header">
  477. <div>
  478. <span class="step-tag s2"><span class="step-num">②</span> 结构化抽取</span>
  479. <span class="step-title">核心主干 · Function Calling 强制 JSON</span>
  480. </div>
  481. <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>
  482. </div>
  483. <div class="pill-row">
  484. <span class="tag llm">LLM 调用</span>
  485. <span class="tag parallel">并行 asyncio.gather + Semaphore(5)</span>
  486. <span class="tag cache">缓存 data/cache/papers/{paper_id}.json + papers.jsonl</span>
  487. </div>
  488. <h4>处理流程</h4>
  489. <ol>
  490. <li><strong>长度判断</strong>:总字符 &gt; 20000 → 按章节分块抽取后合并;否则一次性抽取</li>
  491. <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>
  492. <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>
  493. <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>
  494. </ol>
  495. <details open>
  496. <summary>LLM Prompt · 抽取(核心)</summary>
  497. <pre class="prompt"><span class="tok-sys">[System]</span>
  498. 你是论文结构化抽取助手。严格通过函数调用输出 JSON,
  499. 每个字段含 value/source_sections/confidence。
  500. limitations 必须真实可溯源,严禁臆造。
  501. <span class="tok-user">[User]</span>
  502. 请从以下论文中抽取结构化字段。
  503. 论文标题:<span class="placeholder">{title}</span>
  504. 论文正文:
  505. ## Abstract
  506. <span class="placeholder">{...}</span>
  507. ## Introduction
  508. <span class="placeholder">{...}</span>
  509. 抽取要求:
  510. 1. 必须抽取以下字段:title、task_problem、method、key_contributions、
  511. datasets、metrics、results、limitations、future_work
  512. 2. 每个字段必须是 FieldSource 结构:
  513. {
  514. "value": "字段值(字符串,尽量简洁完整)",
  515. "source_sections": ["原文出处章节标题列表"],
  516. "confidence": 0.0-1.0
  517. }
  518. 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
  519. 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,
  520. 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)
  521. 5. confidence 反映可信度:
  522. - 信息明确且原文直述 → 0.9-1.0
  523. - 需要从原文推断 → 0.6-0.8
  524. - 信息不足或模糊 → 低于 0.6
  525. 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>
  526. </details>
  527. <details>
  528. <summary>LLM Prompt · 针对性补抽(缺失字段)</summary>
  529. <pre class="prompt"><span class="tok-user">[User]</span>
  530. 前一次抽取缺失了以下字段:<span class="placeholder">{", ".join(focus_fields)}</span>。
  531. 请重点从原文中补抽这些字段,其他字段可留空。
  532. <span class="tok-com">{...同上 prompt...}</span></pre>
  533. </details>
  534. <details>
  535. <summary>Function Calling Schema · extract_paper_fields</summary>
  536. <pre>{
  537. "type": "function",
  538. "function": {
  539. "name": "extract_paper_fields",
  540. "description": "从论文中抽取结构化字段,每个字段含 value/source_sections/confidence",
  541. "parameters": {
  542. "type": "object",
  543. "properties": {
  544. "title": { "type": "object", "properties": {
  545. "value": {"type":"string"},
  546. "source_sections": {"type":"array","items":{"type":"string"}},
  547. "confidence": {"type":"number"}
  548. }},
  549. "task_problem": { ... 同上 ... },
  550. "method": { ... },
  551. "key_contributions": { ... },
  552. "datasets": { ... },
  553. "metrics": { ... },
  554. "results": { ... },
  555. "limitations": { ... },
  556. "future_work": { ... }
  557. },
  558. "required": ["title","task_problem","method","key_contributions",
  559. "datasets","metrics","results","limitations","future_work"]
  560. }
  561. }
  562. }</pre>
  563. </details>
  564. <div class="io-grid">
  565. <div class="io-box in">
  566. <div class="label">输入</div>
  567. <pre>{
  568. <span class="tok-str">"paper_path"</span>: <span class="tok-str">"D:/papers/xxx.pdf"</span>,
  569. <span class="tok-str">"parsed_paper"</span>: <span class="tok-com">{ParsedPaper dict}</span>
  570. }</pre>
  571. </div>
  572. <div class="io-box out">
  573. <div class="label">输出:PaperRecord</div>
  574. <pre>{
  575. "path": "...",
  576. "title": {
  577. "value": "Attention Is All You Need",
  578. "source_sections": ["Title"],
  579. "confidence": 1.0
  580. },
  581. "task_problem": {
  582. "value": "序列建模任务...",
  583. "source_sections": ["Introduction"],
  584. "confidence": 0.9
  585. },
  586. "method": {
  587. "value": "Transformer 架构...",
  588. "source_sections": ["3 Method","3.1"],
  589. "confidence": 0.95
  590. },
  591. "limitations": {
  592. "value": "位置编码外推性差...",
  593. "source_sections": ["6 Conclusion"],
  594. "confidence": 0.7
  595. },
  596. <span class="tok-com">...</span>
  597. }</pre>
  598. </div>
  599. </div>
  600. <div class="callout warn">
  601. <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。
  602. </div>
  603. </div>
  604. </section>
  605. <!-- Step 3 -->
  606. <section id="step3">
  607. <div class="card step-card s3">
  608. <div class="step-header">
  609. <div>
  610. <span class="step-tag s3"><span class="step-num">③</span> 向量索引构建</span>
  611. <span class="step-title">BGE-m3 → Chroma + BM25</span>
  612. </div>
  613. <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>
  614. </div>
  615. <div class="pill-row">
  616. <span class="tag local">本地处理</span>
  617. <span class="tag cache">Chroma PersistentClient 持久化</span>
  618. <span class="tag degrade">失败降级 skip_rag</span>
  619. </div>
  620. <h4>处理流程</h4>
  621. <ol>
  622. <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>
  623. <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>
  624. <li><strong>写入 Chroma</strong>:<code class="inline">paper_chunks</code> collection,metadata 含 <code class="inline">paper_path/title/section</code></li>
  625. <li><strong>同步构建 BM25</strong>:倒排索引供关键词检索</li>
  626. </ol>
  627. <div class="io-grid">
  628. <div class="io-box in">
  629. <div class="label">输入</div>
  630. <pre>{
  631. <span class="tok-str">"papers_json"</span>: [<span class="tok-com">{ParsedPaper}</span>, ...]
  632. }</pre>
  633. </div>
  634. <div class="io-box out">
  635. <div class="label">输出</div>
  636. <pre>{
  637. "indexed_count": 12,
  638. "status": "done"
  639. }</pre>
  640. </div>
  641. </div>
  642. <div class="callout">
  643. <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。
  644. </div>
  645. </div>
  646. </section>
  647. <!-- Phase 2 -->
  648. <section id="phase2">
  649. <h2 class="section-title"><span class="num">4</span>Phase 2:Plan(LLM 动态规划)</h2>
  650. <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>
  651. <div class="card">
  652. <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>
  653. <details open>
  654. <summary>LLM Prompt · 动态规划</summary>
  655. <pre class="prompt"><span class="tok-sys">[System]</span>
  656. 你是 Agent 编排器,负责生成可执行的动态计划。只输出 JSON。
  657. <span class="tok-user">[User]</span>
  658. 你是论文创新 Agent 的 Controller。请为以下任务生成动态执行计划。
  659. 任务:分析文件夹 <span class="placeholder">{folder_path}</span> 中的 PDF 论文,
  660. 识别研究空白,生成 2-3 个创新点并设计实验方案。
  661. 可用 Agent:
  662. - PaperParser:扫描文件夹并解析 PDF
  663. - Extractor:单篇论文结构化抽取
  664. - Indexer:向量索引构建
  665. - GapAnalyzer:跨论文 Gap 识别
  666. - InnovationGenerator:创新点生成与筛选
  667. - ExperimentDesigner:实验方案设计
  668. - Integration:整合输出研究提案
  669. 请生成执行计划,输出 JSON 数组,每个元素含:
  670. - step: 步骤序号(整数)
  671. - agent: Agent 名称(上述之一)
  672. - action: 该步骤具体任务描述
  673. - depends_on: 依赖的步骤序号列表
  674. 你可以根据任务自主调整顺序或增减步骤。
  675. 默认链路:解析→抽取→索引→Gap分析→创新生成→实验设计→整合。
  676. 严格只输出 JSON 数组,不要其他文本。</pre>
  677. </details>
  678. <div class="io-grid">
  679. <div class="io-box in">
  680. <div class="label">输入</div>
  681. <pre><span class="tok-str">"D:/papers"</span> <span class="tok-com">或</span>
  682. <span class="tok-str">"已上传的 12 篇论文"</span></pre>
  683. </div>
  684. <div class="io-box out">
  685. <div class="label">输出:计划列表</div>
  686. <pre>[
  687. {"step":4,"agent":"GapAnalyzer",
  688. "action":"跨论文综合分析识别研究空白",
  689. "depends_on":[2]},
  690. {"step":5,"agent":"InnovationGenerator",
  691. "action":"针对 Gap 生成创新点并筛选",
  692. "depends_on":[4]},
  693. <span class="tok-com">...</span>
  694. ]</pre>
  695. </div>
  696. </div>
  697. </div>
  698. </section>
  699. <!-- Phase 3 -->
  700. <section id="phase3">
  701. <h2 class="section-title"><span class="num">5</span>Phase 3:Execute(步骤 4-7 + Reflexion)</h2>
  702. <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>
  703. <div class="callout warn">
  704. <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。
  705. </div>
  706. </section>
  707. <!-- Step 4 -->
  708. <section id="step4">
  709. <div class="card step-card s4">
  710. <div class="step-header">
  711. <div>
  712. <span class="step-tag s4"><span class="step-num">④</span> Gap 识别</span>
  713. <span class="step-title">跨论文综合推理 · 四类 Gap 信号</span>
  714. </div>
  715. <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>
  716. </div>
  717. <div class="pill-row">
  718. <span class="tag llm">LLM 调用</span>
  719. <span class="tag tool">hybrid_search 工具</span>
  720. <span class="tag parallel">RAG 增强 evidence</span>
  721. </div>
  722. <h4>处理流程</h4>
  723. <ol>
  724. <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>
  725. <li><strong>调用 LLM</strong> 做跨论文综合推理,识别四类 Gap</li>
  726. <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>
  727. </ol>
  728. <details open>
  729. <summary>LLM Prompt · Gap 识别</summary>
  730. <pre class="prompt"><span class="tok-sys">[System]</span>
  731. 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。
  732. 只输出严格 JSON,不要输出任何其他内容。
  733. <span class="tok-user">[User]</span>
  734. 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)。
  735. ## 论文对比矩阵
  736. ===== 论文 1: Attention Is All You Need =====
  737. [研究任务/问题] 序列建模任务...
  738. [核心方法] Transformer 架构...
  739. [局限性] 位置编码外推性差...
  740. [未来工作] 扩展到多模态...
  741. ===== 论文 2: BERT =====
  742. [研究任务/问题] ...
  743. [核心方法] ...
  744. [局限性] ...
  745. ## 任务
  746. 请综合分析上述论文,识别以下四类 Gap 信号:
  747. 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
  748. 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
  749. 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
  750. 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
  751. 对于单篇论文场景,重点关注该论文自身的 limitations 与 future_work 字段。
  752. ## 输出要求
  753. 严格输出 JSON:
  754. {
  755. "gaps": [
  756. {
  757. "description": "Gap 的清晰描述",
  758. "gap_type": "repeated_limitation|method_gap|contradictory|unfulfilled_future",
  759. "source_papers": ["来源论文标题1"],
  760. "evidence": "论证依据,引用具体论文的具体字段",
  761. "confidence": 0.0到1.0之间的置信度
  762. }
  763. ]
  764. }
  765. 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>
  766. </details>
  767. <div class="io-grid">
  768. <div class="io-box in">
  769. <div class="label">输入</div>
  770. <pre>{
  771. <span class="tok-str">"papers_records"</span>: [
  772. <span class="tok-com">{PaperRecord 1}</span>,
  773. <span class="tok-com">{PaperRecord 2}</span>,
  774. ...
  775. ]
  776. }</pre>
  777. </div>
  778. <div class="io-box out">
  779. <div class="label">输出</div>
  780. <pre>{
  781. "gaps": [
  782. {
  783. "description": "位置编码外推性普遍较差",
  784. "gap_type": "repeated_limitation",
  785. "source_papers": ["Attention...","BERT..."],
  786. "evidence": "两篇均提及... [原文佐证]\n - 《Attention》6 Conclusion: ...",
  787. "confidence": 0.85
  788. }
  789. ]
  790. }</pre>
  791. </div>
  792. </div>
  793. </div>
  794. </section>
  795. <!-- Reflexion -->
  796. <section id="reflexion">
  797. <div class="card step-card" style="border-left-color: var(--accent-2);">
  798. <div class="step-header">
  799. <div>
  800. <span class="step-tag" style="background:rgba(167,139,250,0.15);color:var(--accent-2);">↻ Reflexion 自校验</span>
  801. <span class="step-title">针对性补抽(非整篇重跑)</span>
  802. </div>
  803. <a class="code-link" href="file:///d:/code/aa-my-idea/agent/paper-agent-trae/backend/orchestrator/reflexion.py#L67">reflexion.py:67</a>
  804. </div>
  805. <div class="pill-row">
  806. <span class="tag llm">LLM 调用</span>
  807. <span class="tag degrade">置信度 &lt; 0.7 触发重试</span>
  808. </div>
  809. <h4>触发条件</h4>
  810. <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>
  811. <details open>
  812. <summary>LLM Prompt · 自校验</summary>
  813. <pre class="prompt"><span class="tok-sys">[System]</span>
  814. 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。
  815. <span class="tok-user">[User]</span>
  816. 请对照上下文检查以下步骤的结果质量。
  817. 步骤:<span class="placeholder">{step_name}</span>
  818. 上下文:
  819. 步骤 2 结果:{抽取记录摘要...}
  820. 结果:
  821. <span class="placeholder">{result_str}</span>
  822. 检查要求:
  823. 1. 评估结果是否准确、完整、无臆造;
  824. 2. 输出 0-1 的置信度(confidence);
  825. 3. 列出发现的问题(issues);
  826. 4. 置信度低于 <span class="placeholder">{CONFIDENCE_THRESHOLD=0.7}</span> 时 needs_retry=true;
  827. 5. 指出需重试的字段或方面(retry_fields)。
  828. 严格以 JSON 输出:
  829. {"confidence": 0.85, "issues": ["问题1"], "needs_retry": false, "retry_fields": []}</pre>
  830. </details>
  831. <h4>重试逻辑</h4>
  832. <ol>
  833. <li><code class="inline">confidence &lt; 0.7</code> → <code class="inline">needs_retry = true</code></li>
  834. <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>
  835. <li>最多重试 <code class="inline">MAX_RETRIES=2</code> 次</li>
  836. <li>仍不达标 → 标记 <code class="inline">needs_human_review</code>,流程继续不阻塞</li>
  837. </ol>
  838. <div class="io-grid">
  839. <div class="io-box in">
  840. <div class="label">输入</div>
  841. <pre>{
  842. <span class="tok-str">"step_name"</span>: <span class="tok-str">"GapAnalyzer"</span>,
  843. <span class="tok-str">"result"</span>: <span class="tok-com">{步骤结果}</span>,
  844. <span class="tok-str">"context"</span>: <span class="tok-com">{依赖步骤结果}</span>
  845. }</pre>
  846. </div>
  847. <div class="io-box out">
  848. <div class="label">输出</div>
  849. <pre>{
  850. "confidence": 0.85,
  851. "issues": ["..."],
  852. "needs_retry": false,
  853. "retry_fields": []
  854. }</pre>
  855. </div>
  856. </div>
  857. </div>
  858. </section>
  859. <!-- Step 5 -->
  860. <section id="step5">
  861. <div class="card step-card s5">
  862. <div class="step-header">
  863. <div>
  864. <span class="step-tag s5"><span class="step-num">⑤</span> 创新点生成</span>
  865. <span class="step-title">四类来源 + 三维评分筛选 + 联网去重</span>
  866. </div>
  867. <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>
  868. </div>
  869. <div class="pill-row">
  870. <span class="tag llm">LLM 调用</span>
  871. <span class="tag tool">web_search 工具(DuckDuckGo)</span>
  872. <span class="tag degrade">联网失败降级 skip_novelty_check</span>
  873. </div>
  874. <h4>处理流程</h4>
  875. <ol>
  876. <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>
  877. <li><strong>规范化 + 排序</strong>:按三维评分之和降序</li>
  878. <li><strong>取 Top 2-3</strong></li>
  879. <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>
  880. </ol>
  881. <details open>
  882. <summary>LLM Prompt · 创新生成</summary>
  883. <pre class="prompt"><span class="tok-sys">[System]</span>
  884. 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。
  885. 只输出严格 JSON,不要输出任何其他内容。
  886. <span class="tok-user">[User]</span>
  887. 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。
  888. ## 用户研究方向约束
  889. 用户的研究方向为:<span class="placeholder">{research_direction}</span>
  890. 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。
  891. ## 研究空白列表
  892. - [repeated_limitation] 位置编码外推性普遍较差(来源: Attention..., BERT...)
  893. - [method_gap] ...
  894. ## 任务
  895. 针对每个 Gap,从以下四类来源生成候选创新方向:
  896. 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
  897. 2. <span class="tok-kw">limitation_fix</span>(局限改进):直接解决被反复提及的局限
  898. 3. <span class="tok-kw">cross_domain</span>(跨域迁移):引入其他领域方法
  899. 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
  900. 对每个候选创新方向,用三维评分(0-10 分):
  901. - novelty(新颖性)
  902. - feasibility(可行性)
  903. - significance(显著性)
  904. ## 输出要求
  905. 严格输出 JSON:
  906. {
  907. "innovations": [
  908. {
  909. "title": "创新点标题",
  910. "idea": "创新思路详细描述",
  911. "source": "method_combination|limitation_fix|cross_domain|new_scenario",
  912. "gap_origin": "对应的 Gap 描述",
  913. "score": {"novelty": 8.0, "feasibility": 7.5, "significance": 8.0},
  914. "supporting_evidence": "论据支撑"
  915. }
  916. ]
  917. }
  918. 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>
  919. </details>
  920. <div class="io-grid">
  921. <div class="io-box in">
  922. <div class="label">输入</div>
  923. <pre>{
  924. <span class="tok-str">"gaps"</span>: [<span class="tok-com">{Gap}</span>, ...],
  925. <span class="tok-str">"research_direction"</span>: <span class="tok-str">"多模态学习"</span>
  926. }</pre>
  927. </div>
  928. <div class="io-box out">
  929. <div class="label">输出</div>
  930. <pre>{
  931. "innovations": [
  932. {
  933. "title": "基于 RoPE 的多模态位置编码",
  934. "idea": "...",
  935. "source": "limitation_fix",
  936. "gap_origin": "位置编码外推性差",
  937. "score": {"novelty":8.0,"feasibility":7.5,"significance":8.5},
  938. "novelty_check": "近 1-2 年无相同工作",
  939. "supporting_evidence": "..."
  940. }
  941. ]
  942. }</pre>
  943. </div>
  944. </div>
  945. </div>
  946. </section>
  947. <!-- Step 6 -->
  948. <section id="step6">
  949. <div class="card step-card s6">
  950. <div class="step-header">
  951. <div>
  952. <span class="step-tag s6"><span class="step-num">⑥</span> 实验方案设计</span>
  953. <span class="step-title">8 要素可执行研究方案</span>
  954. </div>
  955. <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>
  956. </div>
  957. <div class="pill-row">
  958. <span class="tag llm">LLM 调用</span>
  959. </div>
  960. <details open>
  961. <summary>LLM Prompt · 实验设计</summary>
  962. <pre class="prompt"><span class="tok-sys">[System]</span>
  963. 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。
  964. 只输出严格 JSON,不要输出任何其他内容。
  965. <span class="tok-user">[User]</span>
  966. 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。
  967. ## 创新点列表
  968. - 基于 RoPE 的多模态位置编码: ...(来源: limitation_fix,Gap: 位置编码外推性差)
  969. ## 任务
  970. 为每个创新点设计完整、可执行的研究方案,包含:
  971. - <span class="tok-kw">hypothesis</span>(研究假设):清晰可验证的假设
  972. - <span class="tok-kw">datasets</span>(数据集列表):适合该研究的公开数据集
  973. - <span class="tok-kw">baselines</span>(基线方法列表):需要对比的现有方法
  974. - <span class="tok-kw">metrics</span>(评估指标列表):量化评估指标
  975. - <span class="tok-kw">ablation</span>(消融实验设计列表):验证各组件贡献
  976. - <span class="tok-kw">steps</span>(实验步骤列表):可复现的实验流程
  977. - <span class="tok-kw">expected_results</span>(预期结果):预期达到的性能或发现
  978. - <span class="tok-kw">risks</span>(风险):可能的失败点与应对
  979. ## 输出要求
  980. 严格输出 JSON:
  981. {
  982. "experiment_plans": [
  983. {
  984. "innovation_title": "对应的创新点标题",
  985. "hypothesis": "研究假设",
  986. "datasets": ["数据集1","数据集2"],
  987. "baselines": ["基线方法1","基线方法2"],
  988. "metrics": ["指标1","指标2"],
  989. "ablation": ["消融实验1"],
  990. "steps": ["步骤1","步骤2","步骤3"],
  991. "expected_results": "预期结果描述",
  992. "risks": "风险与应对描述"
  993. }
  994. ]
  995. }
  996. 每个创新点对应一条方案,方案须具体可执行。</pre>
  997. </details>
  998. <div class="io-grid">
  999. <div class="io-box in">
  1000. <div class="label">输入</div>
  1001. <pre>{
  1002. <span class="tok-str">"innovations"</span>: [<span class="tok-com">{Innovation}</span>, ...]
  1003. }</pre>
  1004. </div>
  1005. <div class="io-box out">
  1006. <div class="label">输出</div>
  1007. <pre>{
  1008. "experiment_plans": [
  1009. {
  1010. "innovation_title": "基于 RoPE 的多模态位置编码",
  1011. "hypothesis": "RoPE 能提升多模态任务外推性",
  1012. "datasets": ["COCO","VQA"],
  1013. "baselines": ["ALiBi"," vanilla PE"],
  1014. "metrics": ["BLEU","Accuracy"],
  1015. "ablation": ["移除旋转矩阵"],
  1016. "steps": ["..."],
  1017. "expected_results": "...",
  1018. "risks": "..."
  1019. }
  1020. ]
  1021. }</pre>
  1022. </div>
  1023. </div>
  1024. </div>
  1025. </section>
  1026. <!-- Step 7 -->
  1027. <section id="step7">
  1028. <div class="card step-card s7">
  1029. <div class="step-header">
  1030. <div>
  1031. <span class="step-tag s7"><span class="step-num">⑦</span> 整合输出</span>
  1032. <span class="step-title">Markdown 研究提案 · 可追溯</span>
  1033. </div>
  1034. <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>
  1035. </div>
  1036. <div class="pill-row">
  1037. <span class="tag llm">LLM 调用(max_tokens=8192)</span>
  1038. <span class="tag degrade">LLM 失败降级本地格式化</span>
  1039. </div>
  1040. <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>
  1041. <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>
  1042. <pre>{
  1043. "papers_records": [...], <span class="tok-com"># 步骤②所有抽取记录</span>
  1044. "gaps": [...], <span class="tok-com"># 步骤④所有 Gap</span>
  1045. "innovations": [...], <span class="tok-com"># 步骤⑤所有创新点</span>
  1046. "experiment_plans": [...] <span class="tok-com"># 步骤⑥所有实验方案</span>
  1047. }</pre>
  1048. <details open>
  1049. <summary>LLM Prompt · 整合报告</summary>
  1050. <pre class="prompt"><span class="tok-sys">[System]</span>
  1051. 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
  1052. 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。
  1053. <span class="tok-user">[User]</span>
  1054. 请基于以下前序步骤产物,整合为一份结构化研究提案。
  1055. ## 研究主题
  1056. <span class="placeholder">{topic}</span>
  1057. ## 一、论文结构化记录(含 source_sections,可追溯)
  1058. ----- 论文 1: Attention Is All You Need (path=...) -----
  1059. [研究任务/问题] 序列建模... (出处章节: Introduction)
  1060. [核心方法] Transformer... (出处章节: 3 Method; 3.1)
  1061. [局限性] 位置编码外推性差 (出处章节: 6 Conclusion)
  1062. ...
  1063. ## 二、识别到的 Gap 列表
  1064. Gap 1 [repeated_limitation] (置信度=0.85): 位置编码外推性普遍较差
  1065. 来源论文: Attention...; BERT...
  1066. 论证: 两篇均提及...
  1067. ## 三、创新点列表(含三维评分)
  1068. 创新点 1: 基于 RoPE 的多模态位置编码 (来源=limitation_fix)
  1069. 创新思路: ...
  1070. 三维评分: 新颖性 8 / 可行性 7 / 显著性 9
  1071. 新颖性验证: 近 1-2 年无相同工作
  1072. 论据支撑: ...
  1073. ## 四、实验方案列表
  1074. 方案 1 (对应: 基于 RoPE 的多模态位置编码)
  1075. 研究假设: ...
  1076. 数据集: COCO, VQA
  1077. ...
  1078. ## 整合任务
  1079. 1. 背景综述:基于 Gap 分析梳理领域现状与脉络
  1080. 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
  1081. 3. 实验方案:保留每个创新点的完整实验方案要素
  1082. 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)
  1083. 5. Markdown 报告:按指定结构生成完整 Markdown 文本
  1084. ## Markdown 报告结构要求
  1085. # 研究提案:[主题]
  1086. ## 一、背景综述
  1087. ## 二、创新点论证
  1088. ### 创新点 1:[标题]
  1089. - Gap 来源:...
  1090. - 创新思路:...
  1091. - 三维评分:新颖性 X / 可行性 X / 显著性 X
  1092. - 论据支撑:...
  1093. ## 三、实验方案
  1094. ## 四、参考文献溯源
  1095. ## 输出要求
  1096. 严格输出 JSON:
  1097. {
  1098. "background_review": "背景综述文本",
  1099. "innovations": [...],
  1100. "experiment_plans": [...],
  1101. "references": [{"title":"...","path":"...","sections":[...]}],
  1102. "markdown": "完整 Markdown 报告字符串"
  1103. }
  1104. 注意:
  1105. - references 必须可追溯至具体章节(source_sections)
  1106. - 不得编造论文或证据,所有引用须来自上述论文记录
  1107. - 若某字段在输入中缺失,对应位置写"暂无"而非编造</pre>
  1108. </details>
  1109. <div class="io-grid">
  1110. <div class="io-box in">
  1111. <div class="label">输入</div>
  1112. <pre>{
  1113. "papers_records": [...],
  1114. "gaps": [...],
  1115. "innovations": [...],
  1116. "experiment_plans": [...]
  1117. }</pre>
  1118. </div>
  1119. <div class="io-box out">
  1120. <div class="label">输出</div>
  1121. <pre>{
  1122. "background_review": "...",
  1123. "innovations": [...],
  1124. "experiment_plans": [...],
  1125. "references": [
  1126. {"title":"...","path":"...",
  1127. "sections":["3 Method","6 Conclusion"]}
  1128. ],
  1129. "markdown": "# 研究提案:...\n## 一、背景综述\n..."
  1130. }</pre>
  1131. </div>
  1132. </div>
  1133. </div>
  1134. </section>
  1135. <!-- Evaluation -->
  1136. <section id="eval">
  1137. <h2 class="section-title"><span class="num">6</span>评估阶段</h2>
  1138. <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>
  1139. <div class="card">
  1140. <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>
  1141. <ol>
  1142. <li><code class="inline">ToolCallTracker</code> 汇总工具调用成功率 / 总数 / 失败数 / 耗时</li>
  1143. <li><code class="inline">QualityEvaluator</code> 对每个创新点调 LLM 做三维评分校验</li>
  1144. </ol>
  1145. <details open>
  1146. <summary>LLM Prompt · 质量评估</summary>
  1147. <pre class="prompt"><span class="tok-sys">[System]</span>
  1148. 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。
  1149. <span class="tok-user">[User]</span>
  1150. 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)。
  1151. ## 创新点信息
  1152. - 标题:<span class="placeholder">{title}</span>
  1153. - 思路:<span class="placeholder">{idea}</span>
  1154. - 论据支撑:<span class="placeholder">{supporting_evidence}</span>
  1155. - 新颖性检查:<span class="placeholder">{novelty_check}</span>
  1156. ## 评分维度
  1157. - <span class="tok-kw">novelty</span>(新颖性):与现有工作的差异程度
  1158. - <span class="tok-kw">feasibility</span>(可行性):技术实现与资源获取的难度
  1159. - <span class="tok-kw">significance</span>(显著性):学术或应用价值
  1160. ## 输出要求
  1161. 严格输出 JSON:
  1162. {
  1163. "innovation_title": "...",
  1164. "score": {"novelty": 8.0, "feasibility": 7.5, "significance": 8.0},
  1165. "reasoning": "评分理由,简明扼要"
  1166. }</pre>
  1167. </details>
  1168. <h4>输出</h4>
  1169. <pre>{
  1170. "innovations_scores": [...],
  1171. "tool_call_success_rate": 0.89,
  1172. "total_tool_calls": 47,
  1173. "failed_tool_calls": 5,
  1174. "avg_confidence": 0.78,
  1175. "tool_calls": [...],
  1176. "quality_assessment": {...}
  1177. }</pre>
  1178. </div>
  1179. </section>
  1180. <!-- Summary -->
  1181. <section id="summary">
  1182. <h2 class="section-title"><span class="num">7</span>完整 LLM 调用清单</h2>
  1183. <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>
  1184. <table>
  1185. <thead>
  1186. <tr>
  1187. <th>#</th>
  1188. <th>阶段</th>
  1189. <th>调用位置</th>
  1190. <th>Prompt 设计要点</th>
  1191. <th>输出格式</th>
  1192. </tr>
  1193. </thead>
  1194. <tbody>
  1195. <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>
  1196. <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>
  1197. <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>
  1198. <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>
  1199. <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>
  1200. <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>
  1201. <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>
  1202. <tr><td>8</td><td>步骤⑤ Reflexion</td><td>reflexion.py:67</td><td>同上</td><td>JSON confidence</td></tr>
  1203. <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>
  1204. <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>
  1205. <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>
  1206. </tbody>
  1207. </table>
  1208. <h3>工具调用清单</h3>
  1209. <table>
  1210. <thead>
  1211. <tr><th>工具</th><th>步骤</th><th>输入</th><th>输出</th><th>特性</th></tr>
  1212. </thead>
  1213. <tbody>
  1214. <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>
  1215. <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>
  1216. <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>
  1217. <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>
  1218. <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>
  1219. <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>
  1220. <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>
  1221. <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>
  1222. <tr><td><code class="inline">design_experiments</code></td><td>⑥</td><td>innovations</td><td>experiment_plans</td><td>LLM</td></tr>
  1223. <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>
  1224. </tbody>
  1225. </table>
  1226. </section>
  1227. <!-- Principles -->
  1228. <section id="principles">
  1229. <h2 class="section-title"><span class="num">8</span>Prompt 设计的 5 个核心原则</h2>
  1230. <p class="section-desc">贯穿所有 prompt 的设计哲学。</p>
  1231. <div class="card">
  1232. <h4>① System 角色锚定</h4>
  1233. <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>
  1234. <h4>② 结构化输出强制</h4>
  1235. <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>
  1236. <h4>③ 防编造机制</h4>
  1237. <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>
  1238. <h4>④ 上下文注入</h4>
  1239. <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>
  1240. <h4>⑤ 降级兜底</h4>
  1241. <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>
  1242. </div>
  1243. </section>
  1244. <!-- Degradation -->
  1245. <section id="degrade">
  1246. <h2 class="section-title"><span class="num">9</span>降级与容错策略</h2>
  1247. <p class="section-desc">每个步骤都有失败兜底,局部失败不阻塞全局。</p>
  1248. <table>
  1249. <thead>
  1250. <tr><th>失败点</th><th>降级路径</th><th>影响</th></tr>
  1251. </thead>
  1252. <tbody>
  1253. <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>
  1254. <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>
  1255. <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>
  1256. <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>
  1257. <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>
  1258. <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>
  1259. <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>
  1260. <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>
  1261. <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>
  1262. <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>
  1263. </tbody>
  1264. </table>
  1265. <div class="callout ok">
  1266. <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。
  1267. </div>
  1268. </section>
  1269. <div class="footer">
  1270. PaperMind · 论文创新点发现 Agent · 端到端运行流程详解<br>
  1271. 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
  1272. </div>
  1273. </main>
  1274. </div>
  1275. <script>
  1276. // 侧边栏高亮当前 section
  1277. const links = document.querySelectorAll('.sidebar a');
  1278. const sections = document.querySelectorAll('section[id], div[id]');
  1279. const observer = new IntersectionObserver((entries) => {
  1280. entries.forEach(entry => {
  1281. if (entry.isIntersecting) {
  1282. const id = entry.target.id;
  1283. links.forEach(l => {
  1284. l.classList.toggle('active', l.getAttribute('href') === '#' + id);
  1285. });
  1286. }
  1287. });
  1288. }, { rootMargin: '-20% 0px -70% 0px' });
  1289. sections.forEach(s => observer.observe(s));
  1290. </script>
  1291. </body>
  1292. </html>