perf: lazy-import torch and tiktoken in embedding_compute / chat

packages/leann-core/src/leann/chat.py

@@ -10,8 +10,6 @@
 from abc import ABC, abstractmethod
 from typing import Any, Optional, cast
 
-import torch
-
 from .settings import (
     resolve_anthropic_api_key,
     resolve_anthropic_base_url,
@@ -720,6 +718,8 @@ def ask(self, prompt: str, **kwargs) -> str:
         logger.info(f"Generating with HuggingFace model, config: {generation_config}")
 
         # Generate
+        import torch
+
         with torch.no_grad():
             outputs = self.model.generate(**inputs, **generation_config)
 

packages/leann-core/src/leann/embedding_compute.py

@@ -12,11 +12,13 @@
 from typing import Any, Optional, Protocol, cast
 
 import numpy as np
-import tiktoken
-import torch
 
 from .settings import resolve_ollama_host, resolve_openai_api_key, resolve_openai_base_url
 
+# torch and tiktoken are imported lazily inside the functions that use them, so
+# `import leann` (e.g. for MCP search over an existing index, BM25-only flows,
+# or non-embedding utilities) doesn't pull torch's ~1 GB of state into memory.
+
 # Set up logger with proper level
 logger = logging.getLogger(__name__)
 LOG_LEVEL = os.getenv("LEANN_LOG_LEVEL", "WARNING").upper()
@@ -145,6 +147,8 @@ def truncate_to_token_limit(texts: list[str], token_limit: int) -> list[str]:
     if not texts:
         return []
 
+    import tiktoken
+
     # Use tiktoken with cl100k_base encoding
     enc = tiktoken.get_encoding("cl100k_base")
 
@@ -419,6 +423,8 @@ def compute_embeddings_sentence_transformers(
         is_build: Whether this is a build operation (shows progress bar)
         adaptive_optimization: Whether to use adaptive optimization based on batch size
     """
+    import torch
+
     outer_start_time = time.time()
     # Handle empty input
     if not texts: