diff --git a/CLAUDE.md b/CLAUDE.md
index 7d78d96..6d87aa6 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -43,6 +43,7 @@
 - **新增并行点的固定三处**:`common/pool.go` 加池变量(grpool 封装)→ `biz/consts` 加默认大小 → `config.yml` 的 `pool` 段加 `key: 并发度`(缺失或非法时回退默认值)
 - 禁止直接用裸 `go` 启动并行工作负载,一律走 `common` 的池(池清单与默认值见 README 配置说明)
 - **防死锁**:等待链单向「主 → A池 → B池」,被等待池的任务内不得再等待任何池(会饿死 worker);池无 Wait 方法,等待用调用方 `sync.WaitGroup`,任务结果经 buffered channel 回主 goroutine
+- **共享状态安全(内存)**:池内任务并发执行,共享实例(service 单例、model 句柄等)只允许**只读**访问;可变字段必须在**提交池之前**由主 goroutine 一次性预置,任务内禁止写共享字段——Go map 并发写直接 `fatal error: concurrent map writes`,无锁、无降级、不可恢复,只能崩溃重启。需要可变共享状态时按优先级:① 无共享(任务内新建、buffered channel 传递结果)② 锁(`sync.Mutex`/`RWMutex`,锁内只做内存操作,LLM/DB 等 IO 放锁外)③ `sync/atomic`(仅限 int 类标量计数/标志,如 `atomic.AddInt64`,并发计数禁用普通 `++`;复合结构不要用 atomic,指针 CAS 属例外)
 - **裸 `go` 允许的例外**:`go func(){ wg.Wait(); close(ch) }()` 收尾惯用法、SSE 心跳、流式管道(Stream 读写)等长生命周期/非工作负载协程
 
 ## 文档职责（三文档体系）
diff --git a/data/business.db b/data/business.db
index 53b6bd5..828a3b6 100644
Binary files a/data/business.db and b/data/business.db differ
diff --git a/data/chat.db b/data/chat.db
index a277aca..5803fa7 100644
Binary files a/data/chat.db and b/data/chat.db differ
diff --git a/data/system.db b/data/system.db
index 83a9ed4..56b7313 100644
Binary files a/data/system.db and b/data/system.db differ
diff --git a/kb/consts/consts.go b/kb/consts/consts.go
index 9978f64..5eb092e 100644
--- a/kb/consts/consts.go
+++ b/kb/consts/consts.go
@@ -26,7 +26,10 @@ const (
 	MaxReactRounds     = 10   // ReAct 轮次上限
 	ToolResultMaxChars = 1500 // search 工具返回的单条 chunk 截断字数（控制上下文体积）
 
-	ParsePollIntervalSeconds = 5 // 解析/标注任务轮询间隔
+	ParsePollIntervalSeconds = 15 // 解析/标注任务轮询间隔
+
+	DefaultLlmHttpTimeout = 600  // LLM HTTP 客户端超时（秒）：本地模型生成慢，且多请求排队时响应可超分钟级
+	MaxExtractTokens      = 4096 // 知识抽取单次生成的 token 预算（防失控生成烧掉上下文）
 
 	EmbedBatchSize = 16 // 单次向量化请求的文本批量
 
diff --git a/kb/service/chat_service.go b/kb/service/chat_service.go
index b8af69d..549fafe 100644
--- a/kb/service/chat_service.go
+++ b/kb/service/chat_service.go
@@ -33,8 +33,6 @@ import (
 	"github.com/gogf/gf/v2/frame/g"
 )
 
-var httpClient = &http.Client{Timeout: 2 * time.Minute}
-
 // ---------- OpenAI 兼容 HTTP 组件 ----------
 
 type openAIMessage struct {
@@ -76,10 +74,17 @@ type openAIStreamChunk struct {
 type OpenAIChatModel struct {
 	cfg   *entity.ModelConfig
 	tools []*schema.ToolInfo
+	extra map[string]any // 附加请求字段（如 gemma-4 关闭思维链）
 }
 
 func NewOpenAIChatModel(cfg *entity.ModelConfig) *OpenAIChatModel {
-	return &OpenAIChatModel{cfg: cfg}
+	return &OpenAIChatModel{cfg: cfg, extra: map[string]any{}}
+}
+
+// DisableThinking 关闭模型思维链：本地 gemma-4-E4B 偶发长思考会烧光上下文，导致返回空内容
+func (m *OpenAIChatModel) DisableThinking() *OpenAIChatModel {
+	m.extra["thinking"] = false
+	return m
 }
 
 func (m *OpenAIChatModel) Generate(ctx context.Context, input []*schema.Message, opts ...emodel.Option) (*schema.Message, error) {
@@ -88,6 +93,12 @@ func (m *OpenAIChatModel) Generate(ctx context.Context, input []*schema.Message,
 		"messages": buildOpenAIMessages(input),
 		"stream":   false,
 	}
+	for k, v := range m.extra {
+		payload[k] = v
+	}
+	if commonOpts := emodel.GetCommonOptions(nil, opts...); commonOpts.MaxTokens != nil {
+		payload["max_tokens"] = *commonOpts.MaxTokens
+	}
 	m.withTools(payload)
 	body, err := postOpenAI(ctx, m.cfg, m.endpoint("/chat/completions"), payload)
 	if err != nil {
@@ -1174,7 +1185,9 @@ func doOpenAIRequest(ctx context.Context, cfg *entity.ModelConfig, url string, p
 	if cfg.ApiKey != "" {
 		req.Header.Set("Authorization", "Bearer "+cfg.ApiKey)
 	}
-	resp, err := httpClient.Do(req)
+	// 本地模型生成慢且多请求排队时响应可超分钟级，超时取 config.yml chat.timeout（秒）
+	timeout := g.Cfg().MustGet(ctx, "chat.timeout", consts.DefaultLlmHttpTimeout).Int()
+	resp, err := (&http.Client{Timeout: time.Duration(timeout) * time.Second}).Do(req)
 	if err != nil {
 		g.Log().Errorf(ctx, "model call failed: model=%s url=%s err=%v", cfg.ModelName, url, err)
 		return nil, err
diff --git a/kb/service/kg_entity_service.go b/kb/service/kg_entity_service.go
index 7713c28..2d4fadb 100644
--- a/kb/service/kg_entity_service.go
+++ b/kb/service/kg_entity_service.go
@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"strings"
 	"sync"
 
@@ -11,9 +12,9 @@ import (
 	"rag-local/kb/dao"
 	"rag-local/kb/model/entity"
 
+	emodel "github.com/cloudwego/eino/components/model"
 	"github.com/cloudwego/eino/schema"
 	"github.com/gogf/gf/v2/errors/gerror"
-	"github.com/gogf/gf/v2/frame/g"
 )
 
 var KgEntityService = &kgEntityService{}
@@ -48,6 +49,8 @@ func (s *kgEntityService) ExtractDocument(ctx context.Context, datasetId, docume
 	if err != nil {
 		return 0, err
 	}
+	// 池内并发共享 model 实例，只读字段须在此（单 goroutine）预置好
+	model.DisableThinking()
 
 	type extractOut struct {
 		chunkId   int64
@@ -69,7 +72,7 @@ func (s *kgEntityService) ExtractDocument(ctx context.Context, datasetId, docume
 			ch <- extractOut{chunkId: c.Id, err: e}
 		}); err != nil {
 			wg.Done()
-			g.Log().Warningf(ctx, "kg extract chunk %d submit failed: %v", c.Id, err)
+			fmt.Printf("kg extract chunk %d submit failed: %v\n", c.Id, err)
 		}
 	}
 	go func() { wg.Wait(); close(ch) }()
@@ -78,12 +81,12 @@ func (s *kgEntityService) ExtractDocument(ctx context.Context, datasetId, docume
 	for out := range ch {
 		if out.err != nil {
 			failed++
-			g.Log().Warningf(ctx, "kg extract chunk %d failed: %v", out.chunkId, out.err)
+			fmt.Printf("kg extract chunk %d failed: %v\n", out.chunkId, out.err)
 			continue
 		}
 		if err := s.saveExtract(ctx, datasetId, out.chunkId, out.entities, out.relations); err != nil {
 			failed++
-			g.Log().Warningf(ctx, "kg extract chunk %d failed: %v", out.chunkId, err)
+			fmt.Printf("kg extract chunk %d failed: %v\n", out.chunkId, err)
 		}
 	}
 	return failed, nil
@@ -106,13 +109,22 @@ func (s *kgEntityService) callExtract(ctx context.Context, model *OpenAIChatMode
 		{Role: schema.System, Content: "你是知识抽取助手。从文档片段中抽取实体（人名、组织、地名、产品等专有名词）及实体间的关系（动词或介词短语）。只输出 JSON，不要 markdown 代码块或任何解释，格式：{\"entities\":[{\"name\":\"实体名\",\"type\":\"类型\"}],\"relations\":[{\"head\":\"主体\",\"relation\":\"关系\",\"tail\":\"客体\"}]}"},
 		{Role: schema.User, Content: "文档片段：\n" + chunk.Content},
 	}
-	resp, err := model.Generate(ctx, msgs)
+	resp, err := model.Generate(ctx, msgs, emodel.WithMaxTokens(consts.MaxExtractTokens))
 	if err != nil {
 		return nil, nil, err
 	}
 	data, err := parseKgJSON(resp.Content)
 	if err != nil {
-		return nil, nil, err
+		// 模型偶发返回空内容/截断 JSON，重试一次通常正常
+		fmt.Printf("kg extract chunk %d parse failed (%v), retrying\n", chunk.Id, err)
+		resp, err = model.Generate(ctx, msgs, emodel.WithMaxTokens(consts.MaxExtractTokens))
+		if err != nil {
+			return nil, nil, err
+		}
+		data, err = parseKgJSON(resp.Content)
+		if err != nil {
+			return nil, nil, err
+		}
 	}
 	entities := make([]kgEntityItem, 0, len(data.Entities))
 	for _, e := range data.Entities {
@@ -154,7 +166,7 @@ func (s *kgEntityService) saveExtract(ctx context.Context, datasetId, chunkId in
 	return nil
 }
 
-// parseKgJSON 解析模型输出的 JSON，容忍 ```json 代码块包裹
+// parseKgJSON 解析模型输出的 JSON，容忍 ```json 代码块包裹及首个 JSON 对象后的多余内容（模型偶尔续写 ",{...}"）
 func parseKgJSON(content string) (*kgExtractResult, error) {
 	content = strings.TrimSpace(content)
 	content = strings.TrimPrefix(content, "```json")
@@ -163,11 +175,52 @@ func parseKgJSON(content string) (*kgExtractResult, error) {
 	content = strings.TrimSpace(content)
 	var out kgExtractResult
 	if err := json.Unmarshal([]byte(content), &out); err != nil {
-		return nil, gerror.Wrap(err, "解析抽取 JSON 失败")
+		if fixed, ok := extractFirstJSON(content); ok {
+			if err2 := json.Unmarshal([]byte(fixed), &out); err2 == nil {
+				return &out, nil
+			}
+		}
+		return nil, gerror.Wrapf(err, "解析抽取 JSON 失败，模型输出：%q", truncateRunes(content, 200))
 	}
 	return &out, nil
 }
 
+// extractFirstJSON 提取字符串中第一个完整 JSON 对象（平衡花括号，跳过字符串内的花括号）
+func extractFirstJSON(s string) (string, bool) {
+	start := strings.IndexByte(s, '{')
+	if start < 0 {
+		return "", false
+	}
+	depth := 0
+	inStr := false
+	escaped := false
+	for i := start; i < len(s); i++ {
+		ch := s[i]
+		if inStr {
+			if escaped {
+				escaped = false
+			} else if ch == '\\' {
+				escaped = true
+			} else if ch == '"' {
+				inStr = false
+			}
+			continue
+		}
+		switch ch {
+		case '"':
+			inStr = true
+		case '{':
+			depth++
+		case '}':
+			depth--
+			if depth == 0 {
+				return s[start : i+1], true
+			}
+		}
+	}
+	return "", false
+}
+
 func (s *kgEntityService) List(ctx context.Context, datasetId int64, page, pageSize int) ([]*entity.KgEntity, int, error) {
 	return dao.KgEntity.List(ctx, datasetId, page, pageSize)
 }