fix: reduce chunk size to prevent embedding context overflow

nomic-embed-text tokenizes Chinese characters differently from the
estimateTokenCount heuristic (3-5 real tokens per character vs 1
estimated). Large chunks overflow the 8192 token context. Reduce
TargetTokens/MaxTokens to keep chunks within limits.

internal/ai/rag/index.go

@@ -38,9 +38,9 @@ const knowledgeCollectionName = "knowledge_chunks"
 var Index = &index{
 	chunkConfig: ChunkingConfig{
 		Provider:       string(enums.KnowledgeChunkProviderStructured),
-		TargetTokens:   300,
+		TargetTokens:   100,
-		MaxTokens:      400,
+		MaxTokens:      150,
-		OverlapTokens:  40,
+		OverlapTokens:  20,
 		EnableFallback: true,
 	},
 	registry: ragchunk.NewDefaultRegistry(),