fix: better memory usage from 800+ to 500+ (#11796)

Signed-off-by: yihong0618 <zouzou0208@gmail.com>

api/core/rag/datasource/keyword/jieba/jieba_keyword_table_handler.py

@@ -1,18 +1,19 @@
 import re
 from typing import Optional
 
-import jieba
-from jieba.analyse import default_tfidf
-
-from core.rag.datasource.keyword.jieba.stopwords import STOPWORDS
-
 
 class JiebaKeywordTableHandler:
     def __init__(self):
-        default_tfidf.stop_words = STOPWORDS
+        import jieba.analyse
+
+        from core.rag.datasource.keyword.jieba.stopwords import STOPWORDS
+
+        jieba.analyse.default_tfidf.stop_words = STOPWORDS
 
     def extract_keywords(self, text: str, max_keywords_per_chunk: Optional[int] = 10) -> set[str]:
         """Extract keywords with JIEBA tfidf."""
+        import jieba
+
         keywords = jieba.analyse.extract_tags(
             sentence=text,
             topK=max_keywords_per_chunk,
@@ -22,6 +23,8 @@ class JiebaKeywordTableHandler:
 
     def _expand_tokens_with_subtokens(self, tokens: set[str]) -> set[str]:
         """Get subtokens from a list of tokens., filtering for stopwords."""
+        from core.rag.datasource.keyword.jieba.stopwords import STOPWORDS
+
         results = set()
         for token in tokens:
             results.add(token)

api/core/rag/datasource/vdb/oracle/oraclevector.py

@@ -6,10 +6,8 @@ from contextlib import contextmanager
 from typing import Any
 
 import jieba.posseg as pseg
-import nltk
 import numpy
 import oracledb
-from nltk.corpus import stopwords
 from pydantic import BaseModel, model_validator
 
 from configs import dify_config
@@ -202,6 +200,10 @@ class OracleVector(BaseVector):
         return docs
 
     def search_by_full_text(self, query: str, **kwargs: Any) -> list[Document]:
+        # lazy import
+        import nltk
+        from nltk.corpus import stopwords
+
         top_k = kwargs.get("top_k", 5)
         # just not implement fetch by score_threshold now, may be later
         score_threshold = float(kwargs.get("score_threshold") or 0.0)