From 02373831a0ad61b7d939b0ae10902b2783878002 Mon Sep 17 00:00:00 2001
From: Rohit Behera <126186063+r0h1tb@users.noreply.github.com>
Date: Sun, 2 Aug 2026 10:32:51 +0530
Subject: [PATCH] perf(cli): stop importing sentence_transformers for every
 command

Importing ast_rag.cli took 3.65s, of which 3.05s was sentence_transformers
(and transitively torch and transformers), reached via
services/__init__ -> embedding_manager.

Only semantic search needs it. Every other command -- goto, callers, refs,
sig, blocks, cache-stats -- paid the full cost and never used the model.

Moves the import into _get_model(), where it is actually needed, with
TYPE_CHECKING for the annotations.

  import ast_rag.cli   3.65s -> 0.61s

Measured end to end over 15 CLI scenarios against a live index:

  median command latency   3.068s -> 0.558s
  fastest command          2.864s -> 0.487s

Semantic search is unaffected; the model loads on first use and 'query' still
returns its full result set.
---
 ast_rag/services/embedding_manager.py | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/ast_rag/services/embedding_manager.py b/ast_rag/services/embedding_manager.py
index 6e213ae..6c98513 100644
--- a/ast_rag/services/embedding_manager.py
+++ b/ast_rag/services/embedding_manager.py
@@ -21,7 +21,7 @@
 
 import logging
 import uuid
-from typing import Callable, Optional
+from typing import TYPE_CHECKING, Callable, Optional
 
 import numpy as np
 from neo4j import Driver
@@ -35,7 +35,9 @@
     FieldCondition,
     MatchValue,
 )
-from sentence_transformers import SentenceTransformer
+
+if TYPE_CHECKING:  # pragma: no cover - typing only
+    from sentence_transformers import SentenceTransformer
 
 from ast_rag.models import ASTNode, SearchResult, QdrantConfig, EmbeddingConfig, NodeKind
 
@@ -95,7 +97,7 @@ def __init__(
         self._embed_config = embed_config
         self._neo4j_driver = neo4j_driver
         self._client: Optional[QdrantClient] = None
-        self._model: Optional[SentenceTransformer] = None
+        self._model: Optional["SentenceTransformer"] = None
         # Validate hybrid weights if hybrid search is enabled
         if self._embed_config.hybrid_search:
             total = self._embed_config.vector_weight + self._embed_config.keyword_weight
@@ -110,8 +112,16 @@ def __init__(
     # Lazy initialisation
     # ------------------------------------------------------------------
 
-    def _get_model(self) -> SentenceTransformer:
-        """Return local SentenceTransformer.  Not called when remote_url is set."""
+    def _get_model(self) -> "SentenceTransformer":
+        """Return local SentenceTransformer.  Not called when remote_url is set.
+
+        sentence_transformers pulls in torch and transformers and costs ~3s to
+        import. Only semantic search needs it, so it is imported here rather
+        than at module scope -- otherwise every CLI command, including pure
+        graph lookups like `goto` and `callers`, pays that cost.
+        """
+        from sentence_transformers import SentenceTransformer
+
         if self._model is None:
             logger.info(
                 "Loading embedding model locally: %s (device=%s)",