forked from icey1287/SuperMew
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathproject_visualization.html
More file actions
817 lines (714 loc) · 40.6 KB
/
Copy pathproject_visualization.html
File metadata and controls
817 lines (714 loc) · 40.6 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
<!DOCTYPE html>
<html lang="zh-CN">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>SuperMew 项目架构与核心代码解析</title>
<style>
:root {
--primary-color: #4A90E2;
--secondary-color: #F5A623;
--bg-color: #f4f7f6;
--text-color: #333;
--card-bg: #fff;
--border-color: #e0e0e0;
--code-bg: #2d2d2d;
--sidebar-width: 320px;
}
* { box-sizing: border-box; }
body {
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
background-color: var(--bg-color);
color: var(--text-color);
margin: 0;
padding: 0;
display: flex;
flex-direction: column;
height: 100vh;
}
header {
background-color: var(--card-bg);
padding: 20px 40px;
box-shadow: 0 2px 5px rgba(0,0,0,0.05);
display: flex;
align-items: center;
justify-content: space-between;
z-index: 10;
}
header h1 {
margin: 0;
font-size: 24px;
color: var(--primary-color);
}
.container {
display: flex;
flex: 1;
overflow: hidden;
}
/* 侧边栏：项目结构 */
.sidebar {
width: var(--sidebar-width);
background-color: var(--card-bg);
border-right: 1px solid var(--border-color);
overflow-y: auto;
padding: 20px;
flex-shrink: 0;
}
.sidebar h2 {
font-size: 18px;
margin-top: 0;
margin-bottom: 20px;
color: #555;
border-bottom: 2px solid var(--primary-color);
padding-bottom: 10px;
display: inline-block;
}
.tree-view ul {
list-style-type: none;
padding-left: 20px;
margin: 0;
}
.tree-view > ul {
padding-left: 0;
}
.tree-view li {
margin: 8px 0;
cursor: pointer;
line-height: 1.5;
}
.tree-view .folder {
font-weight: bold;
color: #444;
}
.tree-view .folder::before {
content: "📁 ";
}
.tree-view .file {
color: #666;
transition: all 0.2s;
padding: 6px 10px;
border-radius: 6px;
display: flex;
align-items: center;
justify-content: space-between;
}
.tree-view .file span.filename::before {
content: "📄 ";
margin-right: 4px;
}
.tree-view .file:hover {
background-color: #f0f7ff;
color: var(--primary-color);
}
.tree-view .file.active {
background-color: var(--primary-color);
color: white;
font-weight: bold;
box-shadow: 0 2px 4px rgba(74, 144, 226, 0.2);
}
.tree-view .file.active .badge {
background-color: white;
color: var(--primary-color);
}
/* 主内容区 */
.main-content {
flex: 1;
padding: 40px;
overflow-y: auto;
background-color: var(--bg-color);
scroll-behavior: smooth;
}
.card {
background-color: var(--card-bg);
border-radius: 8px;
padding: 30px;
box-shadow: 0 4px 6px rgba(0,0,0,0.05);
margin-bottom: 30px;
border-top: 4px solid var(--primary-color);
}
.card h2 {
margin-top: 0;
color: var(--text-color);
border-bottom: 1px solid #eee;
padding-bottom: 10px;
margin-bottom: 20px;
display: flex;
align-items: center;
gap: 10px;
}
.card h2::before {
content: "📌";
font-size: 1.2em;
}
.card h3 {
color: #444;
margin-top: 25px;
border-left: 3px solid var(--secondary-color);
padding-left: 10px;
}
/* RAG 流程图样式 */
.flow-diagram {
display: flex;
flex-direction: column;
align-items: center;
margin: 30px 0;
padding: 30px 20px;
background: #fafafa;
border: 1px solid #e0e0e0;
border-radius: 8px;
}
.flow-row {
display: flex;
justify-content: center;
align-items: center;
gap: 20px;
margin: 15px 0;
width: 100%;
}
.flow-box {
background: #e1f5fe;
border: 2px solid var(--primary-color);
border-radius: 6px;
padding: 15px 25px;
font-weight: bold;
text-align: center;
min-width: 180px;
box-shadow: 0 2px 4px rgba(0,0,0,0.05);
position: relative;
}
.flow-box small {
display: block;
font-weight: normal;
font-size: 12px;
color: #555;
margin-top: 5px;
}
.flow-box.decision {
background: #fff8e1;
border-color: var(--secondary-color);
border-radius: 50%;
width: 120px;
height: 120px;
display: flex;
flex-direction: column;
align-items: center;
justify-content: center;
min-width: unset;
}
.flow-arrow {
font-size: 24px;
color: #999;
height: 30px;
display: flex;
align-items: center;
justify-content: center;
}
/* 代码展示区 */
pre {
background-color: var(--code-bg);
color: #f8f8f2;
padding: 20px;
border-radius: 8px;
overflow-x: auto;
font-family: "SFMono-Regular", Consolas, "Liberation Mono", Menlo, Courier, monospace;
font-size: 14px;
line-height: 1.5;
box-shadow: inset 0 0 10px rgba(0,0,0,0.5);
border-left: 4px solid var(--primary-color);
}
code {
font-family: inherit;
}
.keyword { color: #ff79c6; }
.string { color: #f1fa8c; }
.comment { color: #6272a4; }
.function { color: #50fa7b; }
.class { color: #8be9fd; font-style: italic; }
.decorator { color: #ffb86c; }
.tag {
display: inline-block;
background: #e8f4fd;
color: var(--primary-color);
padding: 6px 12px;
border-radius: 20px;
font-size: 13px;
margin-right: 8px;
margin-bottom: 8px;
font-weight: 500;
}
/* 动态内容控制 */
.details-section {
display: none;
}
.details-section.active {
display: block;
animation: fadeIn 0.4s ease-out;
}
@keyframes fadeIn {
from { opacity: 0; transform: translateY(15px); }
to { opacity: 1; transform: translateY(0); }
}
.badge {
background-color: #eee;
color: #666;
font-size: 11px;
padding: 2px 6px;
border-radius: 10px;
margin-left: 8px;
font-weight: normal;
}
.tree-view .folder > .badge {
background-color: var(--primary-color);
color: white;
}
ul.feature-list {
line-height: 1.8;
color: #444;
padding-left: 20px;
}
ul.feature-list li {
margin-bottom: 15px;
}
.highlight-text {
background-color: #fff3cd;
padding: 2px 4px;
border-radius: 4px;
font-weight: 500;
}
/* 导航栏 */
.page-nav {
margin-bottom: 30px;
display: flex;
gap: 15px;
flex-wrap: wrap;
}
.page-nav-btn {
background: white;
border: 1px solid var(--border-color);
padding: 10px 20px;
border-radius: 20px;
cursor: pointer;
font-weight: 500;
color: #555;
transition: all 0.2s;
}
.page-nav-btn:hover, .page-nav-btn.active {
background: var(--primary-color);
color: white;
border-color: var(--primary-color);
}
</style>
</head>
<body>
<header>
<h1>🐈 SuperMew - 项目架构解析</h1>
<div style="font-weight: 500; color: #666;">高级 RAG 知识库系统 ｜ FastAPI + LangGraph + Milvus</div>
</header>
<div class="container">
<!-- 左侧项目树 -->
<div class="sidebar">
<h2>项目文件结构</h2>
<div class="tree-view">
<ul>
<li class="folder" onclick="showPage('page-overview')">SuperMew/</li>
<ul>
<li class="file" onclick="showPage('page-overview')">
<span class="filename">main.py</span><span class="badge">入口</span>
</li>
<li class="folder">backend/ <span class="badge">核心代码区</span></li>
<ul>
<li class="file active" id="nav-api" onclick="showPage('page-api')">
<span class="filename">api.py</span><span class="badge">REST路由</span>
</li>
<li class="file" id="nav-rag" onclick="showPage('page-rag')">
<span class="filename">rag_pipeline.py</span><span class="badge">LangGraph</span>
</li>
<li class="file" id="nav-agent" onclick="showPage('page-agent')">
<span class="filename">agent.py</span><span class="badge">流式/Agent</span>
</li>
<li class="file" id="nav-tools" onclick="showPage('page-tools')">
<span class="filename">tools.py</span><span class="badge">工具网关</span>
</li>
<li class="file" id="nav-embedding" onclick="showPage('page-embedding')">
<span class="filename">embedding.py</span><span class="badge">本地BM25</span>
</li>
<li class="file" id="nav-milvus" onclick="showPage('page-milvus')">
<span class="filename">milvus_client.py</span><span class="badge">混合检索库</span>
</li>
<li class="file" id="nav-milvus_writer" onclick="showPage('page-milvus_writer')">
<span class="filename">milvus_writer.py</span><span class="badge">数据写入</span>
</li>
<li class="file" id="nav-doc_loader" onclick="showPage('page-doc_loader')">
<span class="filename">document_loader.py</span><span class="badge">层级切块</span>
</li>
<li class="file" id="nav-parent_store" onclick="showPage('page-parent_store')">
<span class="filename">parent_chunk_store.py</span><span class="badge">本地块存储</span>
</li>
</ul>
<li class="folder">frontend/</li>
<ul>
<li class="file" onclick="showPage('page-frontend')">
<span class="filename">index.html</span><span class="badge">原生UI</span>
</li>
</ul>
</ul>
</ul>
</div>
</div>
<!-- 右侧详情区 -->
<div class="main-content" id="main-scroll">
<!-- ================= PAGE: OVERVIEW (总结) ================= -->
<div id="page-overview" class="details-section active">
<div class="card">
<h2>项目全局总结</h2>
<p>SuperMew 是一个基于 <span class="highlight-text">Self-Reflective RAG (CRAG)</span> 理念构建的高级智能问答知识库系统。它突破了传统简单 RAG（单纯相似度召回然后丢给模型回答）的局限，在召回率、上下文完整度、以及大模型防止幻觉上做了大量的工业级工程设计。</p>
<div style="margin: 25px 0;">
<span class="tag">FastAPI</span>
<span class="tag">LangChain</span>
<span class="tag">LangGraph (状态机编排)</span>
<span class="tag">Milvus (双路向量数据库)</span>
<span class="tag">HuggingFace (本地Embedding)</span>
</div>
<h3>🌟 六大核心工程亮点</h3>
<ul class="feature-list">
<li><strong>LangGraph 自治检索纠错</strong>：如果初步检索相关性差，系统会自动触发退步提问 (Step-back) 或假设性文档 (HyDE) 生成来扩充查询。</li>
<li><strong>双路混合检索引擎</strong>：系统不依赖商业 API，自己在 <code>embedding.py</code> 徒手实现了持久化的 BM25 稀疏向量计算。结合语义密集向量，在 Milvus 中利用 RRF (Reciprocal Rank Fusion) 算法做排名融合，兼顾语义泛化和专有名词精准匹配。</li>
<li><strong>父子层级文档切片 (Parent-Child Chunking)</strong>：为了解决检索碎片化问题，系统对文档执行 L1/L2/L3 三级切片。仅把极小的 L3 存入 Milvus，一旦命中，自动向上溯源拉取大段 L2/L1 上下文喂给大模型。</li>
<li><strong>防幻觉系统级 Prompt</strong>：严格约束 Agent，要求必须基于知识库结果输出，严禁没结果强答或一回合内陷入死循环查库。</li>
<li><strong>高并发流式穿透体系</strong>：RAG 复杂流程非常耗时，系统通过 <code>asyncio.Queue</code> 在 <code>agent.py</code> 中实现了底层引擎状态的实时捕获，在 LLM 思考前就把 “检索进度” 推送给前端。</li>
<li><strong>轻量级无依赖架构</strong>：核心算法除了依赖底层库外，基本采用原生手写（如 BM25 增量状态机、原生 SSE 协议交互等），极其透明可控。</li>
</ul>
</div>
<div class="card">
<h2>LangGraph 核心运转流转图</h2>
<p>展示了当用户提出一个复杂问题时，系统的内部神经元是如何进行图状运转的：</p>
<div class="flow-diagram">
<div class="flow-row">
<div class="flow-box" style="background: #e8f5e9; border-color: #4caf50;">
用户提问
</div>
</div>
<div class="flow-arrow">↓</div>
<div class="flow-row">
<div class="flow-box">
retrieve_initial
<small>Milvus HNSW+BM25 混合检索</small>
</div>
</div>
<div class="flow-arrow">↓</div>
<div class="flow-row">
<div class="flow-box decision">
grade_documents
<small>Grader LLM</small>
<small>相关性评估</small>
</div>
</div>
<div class="flow-row" style="margin: 0;">
<div style="flex:1; text-align:right; padding-right:40px; color: var(--secondary-color); font-weight: bold;">不相关 (No) ↙</div>
<div style="flex:1; padding-left:40px; color: #4caf50; font-weight: bold;">↘ 相关 (Yes)</div>
</div>
<div class="flow-row" style="align-items: flex-start;">
<div style="flex: 1; display: flex; flex-direction: column; align-items: center;">
<div class="flow-box" style="border-color: var(--secondary-color);">
rewrite_question
<small>HyDE 假设 / Step-back 泛化</small>
</div>
<div class="flow-arrow">↓</div>
<div class="flow-box">
retrieve_expanded
<small>用新查询二次搜索</small>
</div>
<div class="flow-arrow">↘</div>
</div>
<div style="flex: 1; display: flex; flex-direction: column; align-items: center; justify-content: center; height: 100%;">
<div class="flow-arrow" style="height: 100px; padding-top: 50px;">↓</div>
</div>
</div>
<div class="flow-row">
<div class="flow-box" style="background: #e8f5e9; border-color: #4caf50; width: 60%;">
generate_answer
<small>Agent 组装父级大上下文回答</small>
</div>
</div>
</div>
</div>
</div>
<!-- ================= PAGE: RAG PIPELINE (重点1) ================= -->
<div id="page-rag" class="details-section">
<div class="card">
<h2>🧠 [重点] AI大脑状态机编排 (backend/rag_pipeline.py)</h2>
<p>普通的 RAG 流程是直线的：“搜索 -> 组装Prompt -> 大模型输出”。但真实的文档搜索往往会失败（比如用户用了同义词、指代不明）。本系统使用 <strong>LangGraph</strong> 建立了一个带有自我纠错机制的循环状态图（Self-Reflective RAG）。</p>
<h3>核心机制 1：文档裁判官 (Grader)</h3>
<p>在初次检索 <code>retrieve_initial</code> 之后，系统不急于让主模型回答，而是先让一个小模型（Grader）去“打分”。如果文档跟问题牛头不对马嘴，直接判定为 <code>no</code>，避免主模型产生幻觉。</p>
<h3>核心机制 2：查询变换 (Query Transformation)</h3>
<p>被 Grader 驳回后，系统进入 <code>rewrite_question_node</code>。这不是简单的换词，而是采用了高阶策略：</p>
<ul class="feature-list">
<li><strong>HyDE (Hypothetical Document Embeddings)</strong>：当问题很概念化时，让模型直接“盲答”写一篇假设性文章，拿这篇文章的向量去搜索。因为答案的行文风格往往和原始文档更接近，相似度极高。</li>
<li><strong>Step-back (退步提问)</strong>：如果问题是具体的“报错日志”，退回一步提问“这个组件的报错排查原理是什么”，从而召回整体结构文档。</li>
</ul>
<h3>核心代码剖析：LangGraph 有向图编译</h3>
<p>这段代码定义了整个 AI 思考的生命周期：</p>
<pre><code><span class="keyword">def</span> <span class="function">build_rag_graph</span>():
<span class="comment"># 1. RAGState 是图里流通的共享内存，包含了问题、文档和轨迹</span>
graph = StateGraph(RAGState)
<span class="comment"># 2. 注册图节点（定义执行函数）</span>
graph.add_node(<span class="string">"retrieve_initial"</span>, retrieve_initial)
graph.add_node(<span class="string">"grade_documents"</span>, grade_documents_node)
graph.add_node(<span class="string">"rewrite_question"</span>, rewrite_question_node)
graph.add_node(<span class="string">"retrieve_expanded"</span>, retrieve_expanded)
<span class="comment"># 3. 画出固定的单行道</span>
graph.set_entry_point(<span class="string">"retrieve_initial"</span>)
graph.add_edge(<span class="string">"retrieve_initial"</span>, <span class="string">"grade_documents"</span>)
<span class="comment"># 4. 【灵魂】条件分叉：根据 grader 评分决定走向</span>
graph.add_conditional_edges(
<span class="string">"grade_documents"</span>,
<span class="keyword">lambda</span> state: state.get(<span class="string">"route"</span>),
{
<span class="string">"generate_answer"</span>: END, <span class="comment"># 裁判给过了：结束图计算，主模型准备输出</span>
<span class="string">"rewrite_question"</span>: <span class="string">"rewrite_question"</span>, <span class="comment"># 没过：去重写问题</span>
},
)
<span class="comment"># 5. 重写完了去二次搜索，完了再结束</span>
graph.add_edge(<span class="string">"rewrite_question"</span>, <span class="string">"retrieve_expanded"</span>)
graph.add_edge(<span class="string">"retrieve_expanded"</span>, END)
<span class="keyword">return</span> graph.compile()</code></pre>
</div>
</div>
<!-- ================= PAGE: EMBEDDING (重点2) ================= -->
<div id="page-embedding" class="details-section">
<div class="card">
<h2>🔢 [重点] 本地增量 BM25 稀疏向量 (backend/embedding.py)</h2>
<p>商业向量模型（如 OpenAI、BAAI）生成的叫做“稠密向量”(Dense)，擅长“语义模糊匹配”。但如果在手册里搜 <span class="highlight-text">"XYZ-9988C 型号"</span>，稠密向量往往全军覆没。系统在此处硬核地在纯 Python 环境下手写了 <strong>BM25 (Sparse) 稀疏向量算法</strong>。</p>
<h3>硬核点 1：不依赖外接服务的纯净状态机</h3>
<p>系统没有外接 ElasticSearch，而是维护了一个 <code>data/bm25_state.json</code>。每次上传/删除新文档，它都会被安全地锁（<code>threading.Lock()</code>）住并重新计算每个词的 <code>doc_freq</code>（逆文档频率）和全库文档总数 <code>total_docs</code>，实现了轻量级的 <strong>增量 IDF</strong>。</p>
<h3>核心代码剖析：TF-IDF 稀疏向量构建</h3>
<p>由于 Milvus <code>SPARSE_INVERTED_INDEX</code> 的要求，我们需要将文本转为一个形如 <code>{词汇ID: 得分}</code> 的字典（其中只包含非零分数，这就叫稀疏）。</p>
<pre><code><span class="keyword">def</span> <span class="function">_sparse_vector_for_text_unlocked</span>(self, text: str):
<span class="comment"># 1. 拆词并统计当前句子的词频 (Term Frequency, TF)</span>
tokens = self.tokenize(text)
doc_len = len(tokens)
tf = Counter(tokens)
sparse_vector = {}
<span class="keyword">for</span> token, freq <span class="keyword">in</span> tf.items():
<span class="comment"># 词典索引，用于给 Milvus 做倒排 ID</span>
idx = self._vocab[token]
<span class="comment"># 2. IDF 计算（逆文档频率）：如果一个词在全库太常见，得分被暴扣</span>
df = self._doc_freq.get(token, <span class="string">0</span>)
<span class="keyword">if</span> df == <span class="string">0</span>:
idf = math.log((n + <span class="string">1</span>) / <span class="string">1</span>)
<span class="keyword">else</span>:
idf = math.log((n - df + <span class="string">0.5</span>) / (df + <span class="string">0.5</span>) + <span class="string">1</span>)
<span class="comment"># 3. 经典 BM25 算法公式</span>
<span class="comment"># numerator: TF 上限抑制 (饱和度参数 k1)</span>
<span class="comment"># denominator: 引入文档长度惩罚 (参数 b)</span>
numerator = freq * (self.k1 + <span class="string">1</span>)
denominator = freq + self.k1 * (<span class="string">1</span> - self.b + self.b * doc_len / avg)
score = idf * numerator / denominator
<span class="keyword">if</span> score > <span class="string">0</span>:
<span class="comment"># { 维度ID: 分数 } 这就是稀疏向量的本质</span>
sparse_vector[idx] = float(score)
<span class="keyword">return</span> sparse_vector, vocab_changed</code></pre>
</div>
</div>
<!-- ================= PAGE: MILVUS CLIENT (重点3) ================= -->
<div id="page-milvus" class="details-section">
<div class="card">
<h2>🗄️ [重点] 异构多路召回与重排 (backend/milvus_client.py)</h2>
<p>这个模块展示了我们在上一步拿到 Dense 和 Sparse 两种向量后，如何将它们存入 Milvus，并在检索时优雅地合并（Fusion）这两种尺度完全不同的分数。</p>
<h3>异构双重索引设计</h3>
<p>建立 Collection 时，不同的向量类型配备了量身定制的武器：</p>
<pre><code><span class="comment"># 给 Dense 向量配 HNSW 图导航索引 (适合语义发散)</span>
index_params.add_index(
field_name=<span class="string">"dense_embedding"</span>, index_type=<span class="string">"HNSW"</span>, metric_type=<span class="string">"IP"</span>
)
<span class="comment"># 给 Sparse 向量配 倒排索引 (倒排索引是关键词检索引擎的核心)</span>
index_params.add_index(
field_name=<span class="string">"sparse_embedding"</span>, index_type=<span class="string">"SPARSE_INVERTED_INDEX"</span>,
metric_type=<span class="string">"IP"</span>, params={<span class="string">"drop_ratio_build"</span>: <span class="string">0.2</span>} <span class="comment"># 压缩长尾垃圾词</span>
)</code></pre>
<h3>核心代码剖析：RRF (Reciprocal Rank Fusion) 算法</h3>
<p>这是混合检索最难的一步：Dense 的距离分数（例如 0.81）和 BM25 的得分（例如 25.4）如同苹果和香蕉，没法加权平均。系统巧妙使用了 <span class="highlight-text">RRF 算法</span>，<strong>彻底抛弃绝对分数，仅使用两边的排名次序来合并</strong>。</p>
<pre><code><span class="keyword">def</span> <span class="function">hybrid_retrieve</span>(self, dense_embedding, sparse_embedding, top_k=<span class="string">5</span>):
<span class="comment"># 1. 构造语义查询：让它多跑一点候选量 (top_k * 2) 参与后面大乱斗</span>
dense_search = AnnSearchRequest(
data=[dense_embedding], anns_field=<span class="string">"dense_embedding"</span>, limit=top_k * <span class="string">2</span>
)
<span class="comment"># 2. 构造关键词纯词频查询</span>
sparse_search = AnnSearchRequest(
data=[sparse_embedding], anns_field=<span class="string">"sparse_embedding"</span>, limit=top_k * <span class="string">2</span>
)
<span class="comment"># 3. 部署 RRF 重排器 (重排分 = 1 / (k参数 + 单路排名))</span>
<span class="comment"># 排名越靠前（单路排名越小），分数越大；如果两路都排名高，分数加成极高</span>
reranker = RRFRanker(k=<span class="string">60</span>)
<span class="comment"># 4. 把并发的两路搜索抛给 Milvus 底层执行，直接获取排好序的金子</span>
results = self.client.hybrid_search(
collection_name=self.collection_name,
reqs=[dense_search, sparse_search], <span class="comment"># 多路并发</span>
ranker=reranker, <span class="comment"># 使用 RRF 统分</span>
limit=top_k
)
<span class="keyword">return</span> results</code></pre>
</div>
</div>
<!-- ================= PAGE: DOC LOADER (重点4) ================= -->
<div id="page-doc_loader" class="details-section">
<div class="card">
<h2>📄 [重点] Parent-Child 文档层级切片 (backend/document_loader.py)</h2>
<p>系统用“三级切块”解决了 RAG 永远的痛：切小了（命中率高但模型看不懂上下文），切大了（一整页去匹配，检索全是干扰项）。</p>
<h3>切块策略设计与存储分离</h3>
<p>文档通过 <code>RecursiveCharacterTextSplitter</code> 滑动窗口变成了三层。但最牛的地方在于：<span class="highlight-text">入库 Milvus 的永远只有最小的碎块（L3），大块（L1/L2）只作为本地 JSON 存储着</span>，只带有血缘 ID（<code>parent_chunk_id</code>）。等极精准的 L3 命中后，系统凭 ID 把沉睡的 L1 大段捞起供模型理解。</p>
<h3>核心代码剖析：血缘级联切分</h3>
<pre><code><span class="keyword">class</span> <span class="class">DocumentLoader</span>:
<span class="keyword">def</span> <span class="function">_split_page_to_three_levels</span>(self, text, base_doc):
<span class="comment"># 【Level 1 - 宏观层】 约 1200 字符的大章大段</span>
level_1_docs = self._splitter_level_1.create_documents([text])
<span class="keyword">for</span> l1_doc <span class="keyword">in</span> level_1_docs:
l1_id = self._build_chunk_id(1) <span class="comment"># 给 L1 颁发唯一身份证</span>
<span class="comment"># 【Level 2 - 中观层】 基于刚才生成的 L1 内容，切出 600 字符中段</span>
level_2_docs = self._splitter_level_2.create_documents([l1_doc.page_content])
<span class="keyword">for</span> l2_doc <span class="keyword">in</span> level_2_docs:
l2_id = self._build_chunk_id(2) <span class="comment"># 给 L2 颁发身份证</span>
<span class="comment"># 【Level 3 - 微观层】 基于 L2 内容，切出 300 字符句子</span>
level_3_docs = self._splitter_level_3.create_documents([l2_doc.page_content])
<span class="keyword">for</span> l3_doc <span class="keyword">in</span> level_3_docs:
<span class="comment"># 最终装箱打包，这个字典里的数据最终流入库里</span>
root_chunks.append({
<span class="string">"text"</span>: l3_doc.page_content,
<span class="string">"chunk_id"</span>: self._build_chunk_id(3), <span class="comment"># 自己的 ID</span>
<span class="string">"parent_chunk_id"</span>: l2_id, <span class="comment"># 【外键】：我老爸是谁</span>
<span class="string">"root_chunk_id"</span>: l1_id, <span class="comment"># 【外键】：我爷爷是谁</span>
<span class="string">"chunk_level"</span>: <span class="string">3</span>, <span class="comment"># 我是第 3 层</span>
})</code></pre>
</div>
</div>
<!-- ================= PAGE: AGENT ================= -->
<div id="page-agent" class="details-section">
<div class="card">
<h2>🤖 Agent 行为约束与流式黑科技 (backend/agent.py)</h2>
<p>负责管理系统的主循环：串联用户记忆、设定大模型行为边界、并提供复杂的 SSE (Server-Sent Events) 双通道流式渲染能力。</p>
<h3>设计亮点：防幻觉的 System Prompt</h3>
<pre><code>system_prompt=(
<span class="string">"You are a cute cat bot that loves to help users. "</span>
<span class="string">"Use search_knowledge_base when users ask document questions. "</span>
<span class="comment"># 核心约束 1：防止疯狂调工具的死循环 (Rate Limit)</span>
<span class="string">"Do not call the same tool repeatedly in one turn. At most one knowledge tool call per turn. "</span>
<span class="comment"># 核心约束 2：强制注意力转移，防止忽略搜索结果</span>
<span class="string">"Once you call search_knowledge_base and receive its result, you MUST immediately produce the Final Answer based on that result. "</span>
<span class="comment"># 核心约束 3：反幻觉 (Anti-Hallucination)</span>
<span class="string">"If the retrieved context is insufficient, answer honestly that you don't know instead of making up facts."</span>
)</code></pre>
<h3>设计亮点：异步队列解决并发通信难题</h3>
<p>在标准的 Python 流程中，Agent 执行 <code>search_knowledge_base</code> 时整个线程会阻塞。为了让前端看到 <code>"🔍正在重写查询..."</code> 这种底层工具发出的进度，我们利用了 <code>asyncio.Queue</code>：</p>
<pre><code><span class="keyword">async def</span> <span class="function">chat_with_agent_stream</span>(user_text, user_id, session_id):
<span class="comment"># 统一输出通道：模型Token和底层工具状态都往这里丢</span>
output_queue = asyncio.Queue()
<span class="comment"># 这个代理类会塞进全局，供深层的 LangGraph 执行时实时把进度塞进队列</span>
<span class="keyword">class</span> <span class="class">_RagStepProxy</span>:
<span class="keyword">def</span> <span class="function">put_nowait</span>(self, step):
output_queue.put_nowait({<span class="string">"type"</span>: <span class="string">"rag_step"</span>, <span class="string">"step"</span>: step})
<span class="comment"># 开启后台独立协程执行大脑推理</span>
<span class="keyword">async def</span> <span class="function">_agent_worker</span>():
<span class="keyword">async for</span> msg <span class="keyword">in</span> agent.astream(...):
<span class="keyword">await</span> output_queue.put({<span class="string">"type"</span>: <span class="string">"content"</span>, <span class="string">"content"</span>: msg})
<span class="keyword">await</span> output_queue.put(<span class="keyword">None</span>) <span class="comment"># 完成信号</span>
asyncio.create_task(_agent_worker())
<span class="comment"># 主循环流式返回给前端</span>
<span class="keyword">while True</span>:
event = <span class="keyword">await</span> output_queue.get()
<span class="keyword">if</span> event <span class="keyword">is None</span>: <span class="keyword">break</span>
<span class="keyword">yield</span> <span class="string">f"data: {json.dumps(event)}\n\n"</span></code></pre>
</div>
</div>
<!-- ================= PAGE: TOOLS ================= -->
<div id="page-tools" class="details-section">
<div class="card">
<h2>🚪 工具网关与状态逃逸 (backend/tools.py)</h2>
<p>这个模块扮演了 <strong>LangChain Agent</strong> 和 <strong>LangGraph RAG 状态机</strong> 之间的桥梁。</p>
<h3>状态逃逸设计</h3>
<p>普通的 LangChain Tool 只能向大模型返回一段 String。但复杂的 RAG 过程产生了大量珍贵的中间日志（<code>rag_trace</code>）。工具在这里提取出干干净净的文本给模型，而把 <code>rag_trace</code> 偷偷扔进了全局对象，等待上面提到的主循环抽走传给前端。</p>
</div>
</div>
<!-- ================= PAGE: MILVUS WRITER ================= -->
<div id="page-milvus_writer" class="details-section">
<div class="card">
<h2>📝 双重向量并行压入 (backend/milvus_writer.py)</h2>
<p>在数据真正落盘到 Milvus 之前调度 Embedding 模型。“一文两算”模式的核心发起地。</p>
</div>
</div>
<!-- ================= PAGE: PARENT STORE ================= -->
<div id="page-parent_store" class="details-section">
<div class="card">
<h2>📦 本地块存储 (backend/parent_chunk_store.py)</h2>
<p>实现本地化的轻量级 JSON 存储库 <code>ParentChunkStore</code>，主要负责在 RAG 返回 L3 碎片数据后，根据 <code>parent_chunk_id</code> 高速取回完整长文本。</p>
</div>
</div>
<!-- ================= PAGE: FRONTEND ================= -->
<div id="page-frontend" class="details-section">
<div class="card">
<h2>🖥️ 前端纯净渲染 (frontend/index.html)</h2>
<p>采用原生 HTML/JS 开发，包含对 SSE 协议 (Server-Sent Events) 的流式解析渲染。</p>
</div>
</div>
<!-- ================= PAGE: API ================= -->
<div id="page-api" class="details-section">
<div class="card">
<h2>🌐 API 核心路由 (backend/api.py)</h2>
<p>FastAPI 的入口文件。通过 <code>StreamingResponse</code> 返回长期存活的流式通讯。</p>
</div>
</div>
</div>
</div>
<script>
// 单页面应用路由系统
function showPage(pageId) {
// 1. 隐藏所有页面区域
document.querySelectorAll('.details-section').forEach(el => {
el.classList.remove('active');
});
// 2. 清除左侧边栏激活状态
document.querySelectorAll('.file').forEach(el => {
el.classList.remove('active');
});
// 3. 显示目标页面
const targetPage = document.getElementById(pageId);
if (targetPage) {
targetPage.classList.add('active');
// 回到顶部
document.getElementById('main-scroll').scrollTop = 0;
} else {
document.getElementById('page-overview').classList.add('active');
}
// 4. 寻找并高亮对应的左侧菜单
const eventTarget = window.event ? window.event.currentTarget : null;
if(eventTarget && eventTarget.classList.contains('file')) {
eventTarget.classList.add('active');
} else {
// 回落机制：通过ID匹配菜单项
const navMap = {
'page-api': 'nav-api',
'page-rag': 'nav-rag',
'page-agent': 'nav-agent',
'page-tools': 'nav-tools',
'page-embedding': 'nav-embedding',
'page-milvus': 'nav-milvus',
'page-milvus_writer': 'nav-milvus_writer',
'page-doc_loader': 'nav-doc_loader',
'page-parent_store': 'nav-parent_store',
'page-frontend': 'nav-frontend' // added frontend mapping
};
if (navMap[pageId]) {
const navEl = document.getElementById(navMap[pageId]);
if (navEl) navEl.classList.add('active');
}
}
}
// 初始化时如果没有任何激活状态，默认激活某些元素
window.onload = function() {
// 确保总览默认显示时没有任何 file 是高亮的
document.querySelectorAll('.file').forEach(el => {
el.classList.remove('active');
});
};
</script>
</body>
</html>