diff --git a/poetry.lock b/poetry.lock
index 33e85d7..b6b7d69 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -1427,13 +1427,13 @@ attrs = ">=19.2.0"
 
 [[package]]
 name = "litellm"
-version = "1.32.1"
+version = "1.23.9"
 description = "Library to easily interface with LLM API providers"
 optional = false
 python-versions = ">=3.8, !=2.7.*, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*, !=3.5.*, !=3.6.*, !=3.7.*"
 files = [
-    {file = "litellm-1.32.1-py3-none-any.whl", hash = "sha256:f93878ec1571c2259c205d8a898d8555e5f70b42425661041b691e1cd87a1820"},
-    {file = "litellm-1.32.1.tar.gz", hash = "sha256:0f45dda4cfa684c2332116862803b8c7c697b2b4ccc5565ae9e9e1aa07d8054a"},
+    {file = "litellm-1.23.9-py3-none-any.whl", hash = "sha256:de69a2880e4ba7e651ffddce1c17cc762d201a2e4b35b5c567740968b035ae83"},
+    {file = "litellm-1.23.9.tar.gz", hash = "sha256:0c1e0e56f4d1d9c8a33da09d6736bde9b21a8ea324db8c05cc3de65c6b4fad7d"},
 ]
 
 [package.dependencies]
@@ -1449,7 +1449,7 @@ tokenizers = "*"
 
 [package.extras]
 extra-proxy = ["streamlit (>=1.29.0,<2.0.0)"]
-proxy = ["PyJWT (>=2.8.0,<3.0.0)", "apscheduler (>=3.10.4,<4.0.0)", "argon2-cffi (>=23.1.0,<24.0.0)", "backoff", "fastapi (>=0.104.1,<0.105.0)", "fastapi-sso (>=0.10.0,<0.11.0)", "gunicorn (>=21.2.0,<22.0.0)", "orjson (>=3.9.7,<4.0.0)", "python-multipart (>=0.0.6,<0.0.7)", "pyyaml (>=6.0.1,<7.0.0)", "rq", "uvicorn (>=0.22.0,<0.23.0)"]
+proxy = ["PyJWT (>=2.8.0,<3.0.0)", "apscheduler (>=3.10.4,<4.0.0)", "backoff", "fastapi (>=0.104.1,<0.105.0)", "fastapi-sso (>=0.10.0,<0.11.0)", "gunicorn (>=21.2.0,<22.0.0)", "orjson (>=3.9.7,<4.0.0)", "python-multipart (>=0.0.6,<0.0.7)", "pyyaml (>=6.0.1,<7.0.0)", "rq", "uvicorn (>=0.22.0,<0.23.0)"]
 
 [[package]]
 name = "llama-cpp-python"
@@ -5257,4 +5257,4 @@ gpu = ["auto-gptq", "autoawq", "optimum"]
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.9,<4.0"
-content-hash = "a6e28b8c86896a5db632a8b8747adcf466264b8b74bf0a120a125b7eb5a25efc"
+content-hash = "996a6c4d0cb8c2a82e2a1d2c9412ea7d26e92c993c29bde94ab07252f70dc7f3"
diff --git a/pyproject.toml b/pyproject.toml
index 4ff407b..46e1203 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -13,7 +13,7 @@ fastapi = "^0.109.0"
 uvicorn = "^0.27.0"
 humanize = "^4.9.0"
 llama-cpp-python = ">=0.2.26, <0.2.56"
-litellm = "^1.23.12"
+litellm = "1.23.9"  # Newer versions than this break macOS builds: https://github.com/BerriAI/litellm/issues/2607
 txtai = {version = "^7.0.0", extras = ["pipeline-llm"]}
 sse-starlette = "^2.0.0"
 llama-index = "^0.10.4"
diff --git a/selfie/__init__.py b/selfie/__init__.py
index de5508d..d204c56 100644
--- a/selfie/__init__.py
+++ b/selfie/__init__.py
@@ -3,6 +3,25 @@
 import logging
 import warnings
 import colorlog
+
+### Preemptive fix based on suggestion in https://github.com/BerriAI/litellm/issues/2607
+# import platform
+# import os
+#
+# os_name = platform.system()
+#
+# if os_name == 'Darwin':  # macOS
+#     cache_dir = os.path.expanduser('~/Library/Caches/TikToken')
+# elif os_name == 'Windows':
+#     cache_dir = os.path.join(os.environ['APPDATA'], 'TikToken', 'Cache')
+# else:  # Assume Linux/Unix
+#     cache_dir = os.path.expanduser('~/TikToken/Cache')
+#
+# # LiteLLM writes to a read-only directory in the built application bundle, try to override it
+# # Source: https://github.com/BerriAI/litellm/pull/1947, with the latest code here: https://github.com/BerriAI/litellm/blob/main/litellm/utils.py
+# os.environ['TIKTOKEN_CACHE_DIR'] = cache_dir
+#
+# # Now we can safely import litellm
 import litellm
 
 # Suppress specific warnings
@@ -18,5 +37,4 @@
 logging.basicConfig(level=logging.INFO)
 logging.getLogger("selfie").setLevel(level=logging.INFO)
 
-
 litellm.drop_params = True