Fix cache persistence bugs

lightrag/operate.py

@@ -859,6 +859,7 @@ async def kg_query(
             .strip()
         )
 
+    if hashing_kv.global_config.get("enable_llm_cache"):
         # Save to cache
         await save_to_cache(
             hashing_kv,
@@ -873,6 +874,7 @@ async def kg_query(
                 cache_type="query",
             ),
         )
+
     return response
 
 
@@ -989,6 +991,7 @@ async def extract_keywords_only(
             "high_level_keywords": hl_keywords,
             "low_level_keywords": ll_keywords,
         }
+        if hashing_kv.global_config.get("enable_llm_cache"):
             await save_to_cache(
                 hashing_kv,
                 CacheData(
@@ -1002,6 +1005,7 @@ async def extract_keywords_only(
                     cache_type="keywords",
                 ),
             )
+
     return hl_keywords, ll_keywords
 
 
@@ -1205,6 +1209,7 @@ async def mix_kg_vector_query(
             .strip()
         )
 
+        if hashing_kv.global_config.get("enable_llm_cache"):
             # 7. Save cache - Only cache after collecting complete response
             await save_to_cache(
                 hashing_kv,
@@ -1973,6 +1978,7 @@ async def naive_query(
             .strip()
         )
 
+    if hashing_kv.global_config.get("enable_llm_cache"):
         # Save to cache
         await save_to_cache(
             hashing_kv,
@@ -2121,6 +2127,7 @@ async def kg_query_with_keywords(
             .strip()
         )
 
+        if hashing_kv.global_config.get("enable_llm_cache"):
             # 7. Save cache - 只有在收集完整响应后才缓存
             await save_to_cache(
                 hashing_kv,

lightrag/utils.py

@@ -731,9 +731,6 @@ async def save_to_cache(hashing_kv, cache_data: CacheData):
         hashing_kv: The key-value storage for caching
         cache_data: The cache data to save
     """
-    if not hashing_kv.global_config.get("enable_llm_cache"):
-        return
-
     # Skip if storage is None or content is a streaming response
     if hashing_kv is None or not cache_data.content:
         return
@@ -776,6 +773,8 @@ async def save_to_cache(hashing_kv, cache_data: CacheData):
         "original_prompt": cache_data.prompt,
     }
 
+    logger.info(f" == LLM cache == saving {cache_data.mode}: {cache_data.args_hash}")
+
     # Only upsert if there's actual new content
     await hashing_kv.upsert({cache_data.mode: mode_cache})
 
@@ -1314,8 +1313,7 @@ async def use_llm_func_with_cache(
 
         res: str = await use_llm_func(input_text, **kwargs)
 
-        # Save to cache
-        logger.info(f" == LLM cache == saving {arg_hash}")
+        if llm_response_cache.global_config.get("enable_llm_cache_for_entity_extract"):
             await save_to_cache(
                 llm_response_cache,
                 CacheData(
@@ -1325,6 +1323,7 @@ async def use_llm_func_with_cache(
                     cache_type=cache_type,
                 ),
             )
+
         return res
 
     # When cache is disabled, directly call LLM