Refa: remove temperature since some LLMs fail to support. (#8981)

### What problem does this PR solve?


### Type of change

- [x] Refactoring

rag/raptor.py

@@ -107,7 +107,7 @@ class RecursiveAbstractiveProcessing4TreeOrganizedRetrieval:
                             ),
                         }
                     ],
-                    {"temperature": 0.3, "max_tokens": self._max_token},
+                    {"max_tokens": self._max_token},
                 )
                 cnt = re.sub(
                     "(······\n由于长度的原因，回答被截断了，要继续吗？|For the content length reason, it stopped, continue?)",

rag/svr/task_executor.py

@@ -103,6 +103,7 @@ MAX_CONCURRENT_CHUNK_BUILDERS = int(os.environ.get('MAX_CONCURRENT_CHUNK_BUILDER
 MAX_CONCURRENT_MINIO = int(os.environ.get('MAX_CONCURRENT_MINIO', '10'))
 task_limiter = trio.Semaphore(MAX_CONCURRENT_TASKS)
 chunk_limiter = trio.CapacityLimiter(MAX_CONCURRENT_CHUNK_BUILDERS)
+embed_limiter = trio.CapacityLimiter(MAX_CONCURRENT_CHUNK_BUILDERS)
 minio_limiter = trio.CapacityLimiter(MAX_CONCURRENT_MINIO)
 kg_limiter = trio.CapacityLimiter(2)
 WORKER_HEARTBEAT_TIMEOUT = int(os.environ.get('WORKER_HEARTBEAT_TIMEOUT', '120'))
@@ -442,7 +443,8 @@ async def embedding(docs, mdl, parser_config=None, callback=None):
 
     cnts_ = np.array([])
     for i in range(0, len(cnts), EMBEDDING_BATCH_SIZE):
-        vts, c = await trio.to_thread.run_sync(lambda: mdl.encode([truncate(c, mdl.max_length-10) for c in cnts[i: i + EMBEDDING_BATCH_SIZE]]))
+        async with embed_limiter:
+            vts, c = await trio.to_thread.run_sync(lambda: mdl.encode([truncate(c, mdl.max_length-10) for c in cnts[i: i + EMBEDDING_BATCH_SIZE]]))
         if len(cnts_) == 0:
             cnts_ = vts
         else: