root/geo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(server): hot-reload config store and reloadable infra clients

Browse Source 
- Replace single Load() with a watching Store that re-reads config files
  (and config.local.yaml) and fans out a ReloadEvent with a per-field diff
  so consumers can decide whether the change is hot-applicable or requires
  a process restart.
- Wrap llm, retrieval, vector store, and object storage clients in
  Reloadable* shells so the bootstrap can swap their underlying impls when
  config changes without re-instantiating handlers.
- Make jwt.Manager and ops TokenIssuer mutable under a lock so secrets and
  TTLs can be rotated live; thread default plan code through a setter on
  the ops AdminUserService.
- Wire ConfigStore through bootstrap and every cmd/main.go, scheduler /
  worker / tenant-api / ops-api start the watcher; services and handlers
  take a config.Provider so they always read current values for things
  like generation.stream_enabled, scheduler dispatch, retrieval, etc.
- Switch shared/config decoding off viper to a Kratos-derived runtime
  package so env placeholders (\${VAR:default}) resolve consistently and
  the same source machinery powers both the loader and the watcher.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
liangxu
2026-05-01 16:01:23 +08:00
parent ce2d8a2907
commit 618399f86d
61 changed files with 3186 additions and 496 deletions
Download Patch File Download Diff File Expand all files Collapse all files
server/internal/tenant/app/knowledge_service.go
+50 -69
View File
@@ -35,24 +35,18 @@ const (
)
type KnowledgeService struct {
pool *pgxpool.Pool
provider retrieval.Provider
store retrieval.VectorStore
objectStorage objectstorage.Client
llmClient llm.Client
logger *zap.Logger
chunkSize int
chunkOverlap int
embeddingBatchSize int
recallLimit int
rerankTopN int
arkBaseURL string
arkAPIKey string
urlMarkdownModel string
httpClient *http.Client
jobs chan knowledgeParseJob
cleanupMu sync.Mutex
cleanupInFlight map[string]struct{}
pool *pgxpool.Pool
provider retrieval.Provider
store retrieval.VectorStore
objectStorage objectstorage.Client
llmClient llm.Client
logger *zap.Logger
cfg knowledgeRuntimeConfig
configProvider runtimeConfigProvider
httpClient *http.Client
jobs chan knowledgeParseJob
cleanupMu sync.Mutex
cleanupInFlight map[string]struct{}
}
type knowledgeParseJob struct {
@@ -169,50 +163,16 @@ func NewKnowledgeService(
queueSize int,
workerCount int,
) *KnowledgeService {
chunkSize := cfg.ChunkSize
if chunkSize <= 0 {
chunkSize = 900
}
chunkOverlap := cfg.ChunkOverlap
if chunkOverlap < 0 {
chunkOverlap = 120
}
batchSize := cfg.EmbeddingBatchSize
if batchSize <= 0 {
batchSize = 16
}
recallLimit := cfg.RecallLimit
if recallLimit <= 0 {
recallLimit = 12
}
rerankTopN := cfg.RerankTopN
if rerankTopN <= 0 {
rerankTopN = 6
}
svc := &KnowledgeService{
pool: pool,
provider: provider,
store: store,
objectStorage: objectStorage,
llmClient: llmClient,
logger: logger,
chunkSize: chunkSize,
chunkOverlap: chunkOverlap,
embeddingBatchSize: batchSize,
recallLimit: recallLimit,
rerankTopN: rerankTopN,
arkBaseURL: strings.TrimSpace(llmCfg.BaseURL),
arkAPIKey: strings.TrimSpace(llmCfg.APIKey),
urlMarkdownModel: strings.TrimSpace(llmCfg.KnowledgeURLModel),
httpClient: &http.Client{Timeout: defaultKnowledgeURLParseTimeout},
cleanupInFlight: make(map[string]struct{}),
}
if svc.urlMarkdownModel == "" {
svc.urlMarkdownModel = strings.TrimSpace(llmCfg.Model)
}
if svc.arkBaseURL == "" {
svc.arkBaseURL = defaultKnowledgeArkBaseURL
pool: pool,
provider: provider,
store: store,
objectStorage: objectStorage,
llmClient: llmClient,
logger: logger,
cfg: knowledgeRuntimeFromConfig(cfg, llmCfg),
httpClient: &http.Client{Timeout: defaultKnowledgeURLParseTimeout},
cleanupInFlight: make(map[string]struct{}),
}
if workerCount > 0 {
@@ -229,6 +189,23 @@ func NewKnowledgeService(
return svc
}
func (s *KnowledgeService) WithConfigProvider(provider runtimeConfigProvider) *KnowledgeService {
s.configProvider = provider
return s
}
func (s *KnowledgeService) runtimeConfig() knowledgeRuntimeConfig {
if s != nil && s.configProvider != nil {
if cfg := s.configProvider.Current(); cfg != nil {
return knowledgeRuntimeFromConfig(cfg.Retrieval, cfg.LLM)
}
}
if s == nil {
return knowledgeRuntimeFromConfig(config.RetrievalConfig{}, config.LLMConfig{})
}
return s.cfg
}
func (s *KnowledgeService) ListGroups(ctx context.Context) ([]KnowledgeGroupResponse, error) {
actor := auth.MustActor(ctx)
rows, err := s.pool.Query(ctx, `
@@ -781,11 +758,12 @@ func (s *KnowledgeService) ResolveContext(
return nil, response.ErrServiceUnavailable(50353, "embedding_failed", "embedding vector is empty")
}
runtimeCfg := s.runtimeConfig()
searchPoints, err := s.store.Search(ctx, retrieval.SearchRequest{
Vector: vectors[0],
TenantID: tenantID,
GroupIDs: searchGroupIDs,
Limit: s.recallLimit,
Limit: runtimeCfg.RecallLimit,
})
if err != nil {
return nil, response.ErrServiceUnavailable(50354, "qdrant_search_failed", err.Error())
@@ -815,9 +793,9 @@ func (s *KnowledgeService) ResolveContext(
documents = append(documents, candidate.Text)
}
reranked, err := s.provider.Rerank(ctx, query, documents, minInt(s.rerankTopN, len(documents)))
reranked, err := s.provider.Rerank(ctx, query, documents, minInt(runtimeCfg.RerankTopN, len(documents)))
if err != nil {
selected := selectKnowledgeCandidateSnippets(candidates, s.rerankTopN)
selected := selectKnowledgeCandidateSnippets(candidates, runtimeCfg.RerankTopN)
selected = s.enrichKnowledgePromptSnippets(ctx, tenantID, selected)
baseContext.Snippets = selected
baseContext.PreciseFacts = s.collectKnowledgeContextPreciseFacts(ctx, tenantID, searchGroupIDs, selected)
@@ -825,7 +803,7 @@ func (s *KnowledgeService) ResolveContext(
return baseContext, nil
}
selected := selectKnowledgeRerankedSnippets(candidates, reranked, s.rerankTopN)
selected := selectKnowledgeRerankedSnippets(candidates, reranked, runtimeCfg.RerankTopN)
selected = s.enrichKnowledgePromptSnippets(ctx, tenantID, selected)
baseContext.Snippets = selected
baseContext.PreciseFacts = s.collectKnowledgeContextPreciseFacts(ctx, tenantID, searchGroupIDs, selected)
@@ -1606,7 +1584,8 @@ func (s *KnowledgeService) validateIngestionDependencies(sourceType string) erro
}
}
if sourceType == "website" {
if strings.TrimSpace(s.arkAPIKey) == "" {
runtimeCfg := s.runtimeConfig()
if strings.TrimSpace(runtimeCfg.ArkAPIKey) == "" {
return response.ErrServiceUnavailable(50358, "knowledge_webpage_parser_unavailable", "ark api key is not configured")
}
if err := s.llmClient.Validate(); err != nil {
@@ -1761,7 +1740,8 @@ func (s *KnowledgeService) executeParseJob(ctx context.Context, job knowledgePar
chunkSource = strings.TrimSpace(parsedContent.Text)
}
chunks := splitKnowledgeTextIntoChunks(chunkSource, s.chunkSize, s.chunkOverlap)
runtimeCfg := s.runtimeConfig()
chunks := splitKnowledgeTextIntoChunks(chunkSource, runtimeCfg.ChunkSize, runtimeCfg.ChunkOverlap)
if len(chunks) == 0 {
err = errors.New("knowledge content is empty after parsing")
_ = s.markKnowledgeIngestionFailed(context.Background(), job.TenantID, job.ItemID, job.ParseTaskID, err)
@@ -2371,9 +2351,10 @@ func (s *KnowledgeService) listActiveKnowledgeItemIDs(
}
func (s *KnowledgeService) embedChunks(ctx context.Context, chunks []string) ([][]float64, error) {
runtimeCfg := s.runtimeConfig()
results := make([][]float64, 0, len(chunks))
for start := 0; start < len(chunks); start += s.embeddingBatchSize {
end := start + s.embeddingBatchSize
for start := 0; start < len(chunks); start += runtimeCfg.EmbeddingBatchSize {
end := start + runtimeCfg.EmbeddingBatchSize
if end > len(chunks) {
end = len(chunks)
}