Add caching for tiktoken encodings

Author: TimPansino

.gitignore

@@ -10,6 +10,9 @@ megalinter-reports/
 # Benchmarks
 .asv/
 
+# LLM Cache Files
+.tiktoken_cache
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

tests/mlmodel_langchain/conftest.py

@@ -16,13 +16,12 @@
 import os
 
 import pytest
-import tiktoken
 from langchain_core.messages.ai import AIMessage
 from langchain_core.messages.tool import ToolMessage
 from langchain_openai import ChatOpenAI, OpenAIEmbeddings
 from testing_support.fixture.event_loop import event_loop as loop
 from testing_support.fixture.vcr import *  # noqa: F403
-from testing_support.fixture.vcr import VCR_IGNORED_HEADERS, VCR_REPLACE_HEADERS
+from testing_support.fixture.vcr import VCR_IGNORED_HEADERS, VCR_REPLACE_HEADERS, VCR_TIKTOKEN_ENCODINGS
 from testing_support.fixtures import collector_agent_registration_fixture, collector_available_fixture
 from testing_support.ml_testing_utils import set_trace_info
 
@@ -77,10 +76,8 @@
     ]
 )
 
-TIKTOKEN_ENCODINGS = ["cl100k_base"]
+VCR_TIKTOKEN_ENCODINGS.extend(["cl100k_base"])
 
-
-# Intercept outgoing requests and log to file for mocking
 EXPECTED_AGENT_RESPONSE = "Hello!"
 EXPECTED_TOOL_OUTPUT = "Hello!"
 
@@ -111,13 +108,6 @@ def embedding_openai_client(openai_clients):
     return embedding_client
 
 
-@pytest.fixture(scope="session")
-def load_tiktoken_encodings():
-    """Load tiktoken encodings before tests run to avoid issues with VCR blocking the network."""
-    for encoding in TIKTOKEN_ENCODINGS:
-        tiktoken.get_encoding(encoding)
-
-
 @pytest.fixture
 def chat_openai_client(openai_clients):
     chat_client, _ = openai_clients

tests/testing_support/fixture/vcr.py

@@ -79,6 +79,7 @@
     raise ImportError("pytest-recording is required to use the vcr fixtures.") from exc
 
 import json
+import os
 from pathlib import Path
 
 import pytest
@@ -88,6 +89,7 @@
 VCR_IGNORED_HEADERS = ["content-length", "traceparent", "tracestate", "user-agent", "x-goog-api-client"]
 VCR_REPLACE_HEADERS = []  # Must be tuples of (header_name, replacement_value)
 VCR_MATCH_ON = ["method", "scheme", "host", "port", "path", "body", "headers", "query"]
+VCR_TIKTOKEN_ENCODINGS = []
 
 
 # === Settings fixtures, required and overridable ===
@@ -217,6 +219,7 @@ def vcr_config(
     vcr_match_on,
     vcr_before_record_request,
     vcr_before_record_response,
+    vcr_cache_tiktoken_encodings,
 ):
     """
     Combines the overridable settings fixtures into VCR.py's final configuration.
@@ -297,3 +300,22 @@ def pytest_collection_modifyitems(items):
     """
     for item in items:
         item.add_marker(pytest.mark.vcr)
+
+
+@pytest.fixture
+def vcr_cache_tiktoken_encodings(monkeypatch):
+    """Cache the tiktoken encodings before enabling VCR which blocks network access."""
+    try:
+        import tiktoken
+    except ImportError:
+        return  # tiktoken is not installed, skip caching
+
+    # Set up temporary cache dir
+    tox_env_dir = os.environ.get("TOX_ENV_DIR", None) or Path.cwd()
+    cache_dir = Path(tox_env_dir) / ".tiktoken_cache"
+    monkeypatch.setenv("TIKTOKEN_CACHE_DIR", str(cache_dir))
+    cache_dir.mkdir(parents=True, exist_ok=True)
+
+    # Pre-fetch encodings used in tests
+    for encoding in VCR_TIKTOKEN_ENCODINGS:
+        tiktoken.get_encoding(encoding)