feat(rag): add hybrid knowledge rebuild API and script

Add confirm-gated rebuild-hybrid endpoint that drops biz_hybrid, clears
api_document and L0, then force-imports knowledge_base markdown into the
dense+BM25 store. Include PowerShell runner and ops README.
This commit is contained in:
zhuyongxin
2026-07-27 18:55:00 +08:00
parent f035538531
commit 5c369f3b6c
6 changed files with 407 additions and 3 deletions
@@ -0,0 +1,97 @@
# 重建 hybrid 知识库(dense + BM25)
面向当前 `knowledge_base/` 目录文档,清空并重建 `biz_hybrid` collection。
## 前提
1. 应用已启动(默认 `http://localhost:9900`)
2. `MILVUS_TOKEN` 等连接配置可用
3. `application.yml` 已配置:
```yaml
milvus:
collection: biz_hybrid
retrieval:
search:
mode: hybrid
knowledge:
base-path: knowledge_base/
```
## 一键脚本
在项目根目录执行:
```powershell
.\scripts\rebuild-hybrid-knowledge.ps1 -Confirm REBUILD
```
指定服务地址:
```powershell
.\scripts\rebuild-hybrid-knowledge.ps1 -BaseUrl http://127.0.0.1:9900 -Confirm REBUILD
```
## 脚本会做什么
| 步骤 | 动作 |
|---|---|
| 1 | 检查 `/milvus/health` |
| 2 | 打印重建前 `/api/knowledge/stats` |
| 3 | `POST /api/knowledge/rebuild-hybrid?confirm=REBUILD` |
| 4 | 打印重建后 stats |
服务端 `rebuild-hybrid` 内部顺序:
1. **Drop + recreate** Milvus collection(`milvus.collection`,默认 `biz_hybrid`)
2. **清空** MySQL `api_document`
3. **清空** 内存 L0 索引
4. **扫描** `knowledge_base/**/*.md` 并 `force` 全量导入
- 写 MySQL 元数据
- 切片
- 写 dense 向量 + BM25 `search_text`
- 更新 L0
## 不会做什么
- **不会**删除旧 collection `biz`(需你在 Zilliz 控制台自行决定是否删)
- **不会**动 `knowledge_base/` 源文件
- **不会**在未传 `confirm=REBUILD` 时执行
## 手动 curl 等价命令
```bash
# 重建(危险)
curl -X POST "http://localhost:9900/api/knowledge/rebuild-hybrid?confirm=REBUILD"
# 仅增量/强制导入(不 drop collection)
curl -X POST "http://localhost:9900/api/knowledge/init?force=true"
# 统计
curl "http://localhost:9900/api/knowledge/stats"
```
## 成功判据
响应中大致应有:
```json
{
"success": true,
"collection": "biz_hybrid",
"inserted": <大于0>,
"failed": 0,
"milvus": { "recreated": true, "loaded": true }
}
```
然后用一条知识库里真实存在的术语/故障词走 `lookup_knowledge` 或 chat 验证 hybrid 命中。
## 失败排查
| 现象 | 可能原因 |
|---|---|
| connect / token 错误 | `MILVUS_TOKEN`、host、database |
| BM25 / analyzer 相关报错 | 云端 Milvus/Zilliz 版本不支持 BM25 Function |
| inserted=0 | `knowledge_base` 路径不对,或 md 缺 frontmatter/title |
| failed>0 | 看响应 `details` 与应用日志 |
+157
View File
@@ -0,0 +1,157 @@
<#
.SYNOPSIS
清空 hybrid 知识库(Milvus biz_hybrid + MySQL api_document + L0),并从 knowledge_base 全量重建。
.DESCRIPTION
对应 dense+BM25 混合检索上线后的数据迁移步骤:
1) 检查服务健康
2) 可选:查看当前 /api/knowledge/stats
3) POST /api/knowledge/rebuild-hybrid?confirm=REBUILD
- drop + recreate milvus.collection(默认 biz_hybrid)
- 清空 MySQL api_document
- 清空内存 L0 索引
- 扫描 knowledge_base/**/*.md 强制导入并写入 dense+BM25
不会删除旧的 legacy collection `biz`(知识主路径已不再使用它)。
.PARAMETER BaseUrl
服务根地址,默认 http://localhost:9900
.PARAMETER Confirm
必须为 REBUILD 才会真正执行(防止误触)
.PARAMETER SkipStats
跳过重建前后的 stats 查询
.EXAMPLE
# 先启动 Spring Boot,再执行:
.\scripts\rebuild-hybrid-knowledge.ps1 -Confirm REBUILD
.EXAMPLE
.\scripts\rebuild-hybrid-knowledge.ps1 -BaseUrl http://127.0.0.1:9900 -Confirm REBUILD
#>
[CmdletBinding()]
param(
[string]$BaseUrl = "http://localhost:9900",
[ValidateSet("REBUILD")]
[Parameter(Mandatory = $true)]
[string]$Confirm,
[switch]$SkipStats
)
$ErrorActionPreference = "Stop"
function Write-Step([string]$Message) {
Write-Host ""
Write-Host "==> $Message" -ForegroundColor Cyan
}
function Invoke-Json {
param(
[Parameter(Mandatory = $true)][string]$Method,
[Parameter(Mandatory = $true)][string]$Url,
[int]$TimeoutSec = 3600
)
try {
$resp = Invoke-WebRequest -Method $Method -Uri $Url -TimeoutSec $TimeoutSec
$body = $resp.Content
if ([string]::IsNullOrWhiteSpace($body)) {
return @{ statusCode = [int]$resp.StatusCode; body = $null }
}
return @{
statusCode = [int]$resp.StatusCode
body = $body | ConvertFrom-Json
}
}
catch {
$status = $null
$raw = $null
if ($_.Exception.Response) {
$status = [int]$_.Exception.Response.StatusCode
try {
$stream = $_.Exception.Response.GetResponseStream()
$reader = New-Object System.IO.StreamReader($stream)
$raw = $reader.ReadToEnd()
}
catch { }
}
throw "HTTP $Method $Url failed (status=$status): $($_.Exception.Message)`n$raw"
}
}
$BaseUrl = $BaseUrl.TrimEnd("/")
Write-Host "Hybrid knowledge rebuild" -ForegroundColor Green
Write-Host " BaseUrl : $BaseUrl"
Write-Host " Confirm : $Confirm"
Write-Host " Source : knowledge_base/ (server-side knowledge.base-path)"
Write-Host ""
Write-Host "This will DESTROY data in:" -ForegroundColor Yellow
Write-Host " - Milvus collection milvus.collection (default: biz_hybrid)"
Write-Host " - MySQL table api_document"
Write-Host " - In-memory L0 knowledge index"
Write-Host "Then re-import all markdown under knowledge_base."
Write-Host ""
# 1) health
Write-Step "Check service health"
try {
$health = Invoke-Json -Method GET -Url "$BaseUrl/milvus/health" -TimeoutSec 30
Write-Host (" milvus health status={0}" -f $health.statusCode)
if ($health.body) {
$health.body | ConvertTo-Json -Depth 6 | Write-Host
}
}
catch {
Write-Host " WARN: /milvus/health failed: $_" -ForegroundColor Yellow
Write-Host " Continue if app is up but milvus health endpoint has issues."
}
# 2) stats before
if (-not $SkipStats) {
Write-Step "Knowledge stats (before)"
try {
$before = Invoke-Json -Method GET -Url "$BaseUrl/api/knowledge/stats" -TimeoutSec 30
$before.body | ConvertTo-Json -Depth 6 | Write-Host
}
catch {
Write-Host " WARN: stats before failed: $_" -ForegroundColor Yellow
}
}
# 3) rebuild
Write-Step "POST /api/knowledge/rebuild-hybrid?confirm=REBUILD"
$rebuildUrl = "$BaseUrl/api/knowledge/rebuild-hybrid?confirm=$Confirm"
$rebuild = Invoke-Json -Method POST -Url $rebuildUrl -TimeoutSec 7200
Write-Host (" HTTP {0}" -f $rebuild.statusCode)
$rebuild.body | ConvertTo-Json -Depth 12 | Write-Host
if (-not $rebuild.body.success) {
Write-Host ""
Write-Host "Rebuild reported failure. Inspect details above." -ForegroundColor Red
exit 2
}
# 4) stats after
if (-not $SkipStats) {
Write-Step "Knowledge stats (after)"
try {
$after = Invoke-Json -Method GET -Url "$BaseUrl/api/knowledge/stats" -TimeoutSec 30
$after.body | ConvertTo-Json -Depth 6 | Write-Host
}
catch {
Write-Host " WARN: stats after failed: $_" -ForegroundColor Yellow
}
}
Write-Host ""
Write-Host "Done." -ForegroundColor Green
Write-Host "Next:"
Write-Host " 1) Ensure application.yml has:"
Write-Host " milvus.collection: biz_hybrid"
Write-Host " retrieval.search.mode: hybrid"
Write-Host " 2) Smoke test lookup_knowledge / chat with a known doc query"
Write-Host " 3) Optional: drop legacy collection 'biz' manually in Zilliz console if no longer needed"
exit 0
@@ -63,6 +63,59 @@ public class KnowledgeBaseController {
}
}
/**
* 清空 hybrid collection + MySQL api_document + L0 内存索引,
* 再从 knowledge_base 全量重建 dense+BM25 索引。
*
* <p>危险操作:会删除 {@code milvus.collection}(默认 biz_hybrid)与文档元数据表数据。
* 需要显式 confirm=REBUILD。</p>
*/
@PostMapping("/rebuild-hybrid")
public ResponseEntity<?> rebuildHybrid(
@RequestParam(defaultValue = "") String confirm) {
if (!"REBUILD".equals(confirm)) {
Map<String, Object> rejected = new HashMap<>();
rejected.put("success", false);
rejected.put("message", "拒绝执行:请传 confirm=REBUILD 以确认清空并重建");
rejected.put("hint", "POST /api/knowledge/rebuild-hybrid?confirm=REBUILD");
return ResponseEntity.badRequest().body(rejected);
}
logger.warn("收到 hybrid 知识库全量重建请求 confirm={}", confirm);
try {
KnowledgeBaseInitService.RebuildResult result = initService.rebuildHybridFromKnowledgeBase();
Map<String, Object> response = new HashMap<>();
response.put("success", result.isSuccess());
response.put("message", result.isSuccess()
? "hybrid 知识库重建完成"
: "hybrid 知识库重建结束,但存在失败项");
response.put("collection", result.getCollection());
response.put("basePath", result.getBasePath());
response.put("milvus", result.getMilvus());
response.put("mysqlDocumentsBefore", result.getMysqlDocumentsBefore());
response.put("mysqlDocumentsAfterClear", result.getMysqlDocumentsAfterClear());
response.put("mysqlDocumentsAfterInit", result.getMysqlDocumentsAfterInit());
response.put("l0IndexSizeAfterClear", result.getL0IndexSizeAfterClear());
response.put("l0IndexSizeAfterInit", result.getL0IndexSizeAfterInit());
if (result.getInit() != null) {
response.put("scanned", result.getInit().getScanned());
response.put("skipped", result.getInit().getSkipped());
response.put("inserted", result.getInit().getInserted());
response.put("failed", result.getInit().getFailed());
response.put("details", result.getInit().getDetails());
}
return result.isSuccess()
? ResponseEntity.ok(response)
: ResponseEntity.status(500).body(response);
} catch (Exception e) {
logger.error("hybrid 知识库重建失败", e);
Map<String, Object> response = new HashMap<>();
response.put("success", false);
response.put("message", "重建失败: " + e.getMessage());
return ResponseEntity.internalServerError().body(response);
}
}
/**
* 查询知识库统计信息
*
@@ -52,6 +52,42 @@ public class KnowledgeBaseInitService {
@Autowired
private KnowledgeIndexService knowledgeIndexService;
@Autowired
private com.superbiz.agent.service.milvus.MilvusHybridKnowledgeStore hybridKnowledgeStore;
/**
* Drop hybrid collection, clear MySQL api_document + L0 memory index,
* then force-import all markdown under knowledge.base-path into biz_hybrid.
*/
@Transactional(rollbackFor = Exception.class)
public RebuildResult rebuildHybridFromKnowledgeBase() {
logger.info("Starting hybrid knowledge rebuild from {}", knowledgeBasePath);
RebuildResult rebuild = new RebuildResult();
rebuild.setBasePath(knowledgeBasePath);
rebuild.setCollection(hybridKnowledgeStore.collectionName());
long mysqlBefore = apiDocumentRepository.count();
rebuild.setMysqlDocumentsBefore(mysqlBefore);
Map<String, Object> milvus = hybridKnowledgeStore.dropAndRecreateCollection();
rebuild.setMilvus(milvus);
apiDocumentRepository.deleteAll();
apiDocumentRepository.flush();
knowledgeIndexService.clearIndex();
rebuild.setMysqlDocumentsAfterClear(apiDocumentRepository.count());
rebuild.setL0IndexSizeAfterClear(knowledgeIndexService.getIndexSize());
InitResult init = initializeKnowledgeBase(true);
rebuild.setInit(init);
rebuild.setL0IndexSizeAfterInit(knowledgeIndexService.getIndexSize());
rebuild.setMysqlDocumentsAfterInit(apiDocumentRepository.count());
rebuild.setSuccess(init.getFailed() == 0 && init.getInserted() > 0);
logger.info("Hybrid knowledge rebuild finished: success={}, inserted={}, failed={}",
rebuild.isSuccess(), init.getInserted(), init.getFailed());
return rebuild;
}
/**
* 初始化知识库
*
@@ -131,16 +167,16 @@ public class KnowledgeBaseInitService {
List<DocumentChunk> chunks = documentChunkService.chunkDocument(body, relativePath);
logger.debug("文档分块完成: {} -> {} 个 chunk", relativePath, chunks.size());
// 上传到 Milvus
// 上传到 Milvus hybrid collection(dense + BM25 search_text)
try {
vectorIndexService.indexDocumentChunks(document.getDocId(), chunks, category);
vectorIndexService.indexDocumentChunks(document.getDocId(), chunks, category, frontmatter);
document.setStatus("INDEXED");
document.setChunkCount(chunks.size());
document.setIndexedAt(LocalDateTime.now());
apiDocumentRepository.save(document);
logger.info("文档已索引到 Milvus: {} (docId={}, chunks={})",
logger.info("文档已索引到 Milvus hybrid: {} (docId={}, chunks={})",
title, document.getDocId(), chunks.size());
} catch (Exception e) {
logger.error("上传到 Milvus 失败: {}", relativePath, e);
@@ -161,6 +197,7 @@ public class KnowledgeBaseInitService {
.keywords(keywords)
.summary(summary)
.category(category)
.kbScope(frontmatter.getKbScope())
.build();
knowledgeIndexService.addToIndex(entry);
@@ -344,4 +381,21 @@ public class KnowledgeBaseInitService {
private long totalVectors; // Milvus 中的向量总数
private Map<String, Long> categoryCount; // 按分类统计
}
/**
* Full hybrid rebuild result.
*/
@Data
public static class RebuildResult {
private boolean success;
private String basePath;
private String collection;
private long mysqlDocumentsBefore;
private long mysqlDocumentsAfterClear;
private long mysqlDocumentsAfterInit;
private int l0IndexSizeAfterClear;
private int l0IndexSizeAfterInit;
private Map<String, Object> milvus;
private InitResult init;
}
}
@@ -296,6 +296,12 @@ public class KnowledgeIndexService {
log.debug("文档已从 L0 索引移除: {}", filePath);
}
/** Clear in-memory L0 entries (used by knowledge rebuild). */
public void clearIndex() {
knowledgeIndex.clear();
log.info("L0 knowledge index cleared");
}
public int getIndexSize() {
return knowledgeIndex.size();
}
@@ -12,8 +12,10 @@ import io.milvus.v2.common.DataType;
import io.milvus.v2.common.IndexParam;
import io.milvus.v2.service.collection.request.AddFieldReq;
import io.milvus.v2.service.collection.request.CreateCollectionReq;
import io.milvus.v2.service.collection.request.DropCollectionReq;
import io.milvus.v2.service.collection.request.HasCollectionReq;
import io.milvus.v2.service.collection.request.LoadCollectionReq;
import io.milvus.v2.service.collection.request.ReleaseCollectionReq;
import io.milvus.v2.service.index.request.CreateIndexReq;
import io.milvus.v2.service.vector.request.AnnSearchReq;
import io.milvus.v2.service.vector.request.DeleteReq;
@@ -141,6 +143,41 @@ public class MilvusHybridKnowledgeStore {
.build());
}
/**
* Drop the hybrid collection (if present) and recreate empty dense+BM25 schema.
* Used by knowledge rebuild scripts. Does not touch legacy {@code biz} collection.
*/
public synchronized Map<String, Object> dropAndRecreateCollection() {
Map<String, Object> result = new LinkedHashMap<>();
result.put("collection", collectionName);
MilvusClientV2 milvusClient = client();
Boolean exists = milvusClient.hasCollection(HasCollectionReq.builder()
.collectionName(collectionName)
.build());
result.put("existedBefore", Boolean.TRUE.equals(exists));
if (Boolean.TRUE.equals(exists)) {
try {
milvusClient.releaseCollection(ReleaseCollectionReq.builder()
.collectionName(collectionName)
.build());
} catch (Exception e) {
log.warn("Release collection before drop failed (continuing): {}", e.getMessage());
}
milvusClient.dropCollection(DropCollectionReq.builder()
.collectionName(collectionName)
.build());
log.info("Dropped hybrid collection '{}'", collectionName);
result.put("dropped", true);
} else {
result.put("dropped", false);
}
ensureCollection(milvusClient);
loadCollection(milvusClient);
result.put("recreated", true);
result.put("loaded", true);
return result;
}
public List<VectorSearchService.SearchResult> searchDense(String queryEmbeddingText,
List<Float> queryVector,
int topK,