From 654664795f2f01c4691fd9e89483060ac9cac9a0 Mon Sep 17 00:00:00 2001
From: Petie Clark <petie@clarktechcustoms.com>
Date: Thu, 19 Mar 2026 17:46:30 -0400
Subject: [PATCH] fix(redis-semantic-cache): support custom embedding api_base
 and api_key

RedisSemanticCache previously hardcoded the embedding call to always use
the default OpenAI endpoint, making it impossible to use self-hosted or
alternative embedding providers (e.g. LM Studio, Ollama, Azure).

Changes:
- Add embedding_api_base and embedding_api_key params to RedisSemanticCache.__init__
- Store on self and pass through to litellm.embedding() in _get_embedding()
- Expose as redis_semantic_cache_embedding_api_base and
  redis_semantic_cache_embedding_api_key on the Cache factory class

Backward compatible: both params default to None; existing deployments
using text-embedding-ada-002 via OpenAI are unaffected.

Fixes the startup crash where CustomTextVectorizer.__init__ calls
_get_embedding() synchronously and fails with a 404/auth error when the
embedding model is not accessible at the default OpenAI endpoint.
---
 litellm/caching/caching.py              |  4 ++++
 litellm/caching/redis_semantic_cache.py | 19 ++++++++++++++-----
 2 files changed, 18 insertions(+), 5 deletions(-)

diff --git a/litellm/caching/caching.py b/litellm/caching/caching.py
index 406a4f8c98a..5eca831ab59 100644
--- a/litellm/caching/caching.py
+++ b/litellm/caching/caching.py
@@ -99,6 +99,8 @@ class Cache:
         gcs_path_service_account: Optional[str] = None,
         gcs_path: Optional[str] = None,
         redis_semantic_cache_embedding_model: str = "text-embedding-ada-002",
+        redis_semantic_cache_embedding_api_base: Optional[str] = None,
+        redis_semantic_cache_embedding_api_key: Optional[str] = None,
         redis_semantic_cache_index_name: Optional[str] = None,
         redis_flush_size: Optional[int] = None,
         redis_startup_nodes: Optional[List] = None,
@@ -205,6 +207,8 @@ class Cache:
                 password=password,
                 similarity_threshold=similarity_threshold,
                 embedding_model=redis_semantic_cache_embedding_model,
+                embedding_api_base=redis_semantic_cache_embedding_api_base,
+                embedding_api_key=redis_semantic_cache_embedding_api_key,
                 index_name=redis_semantic_cache_index_name,
                 **kwargs,
             )
diff --git a/litellm/caching/redis_semantic_cache.py b/litellm/caching/redis_semantic_cache.py
index c76f27377d8..56c9bd61621 100644
--- a/litellm/caching/redis_semantic_cache.py
+++ b/litellm/caching/redis_semantic_cache.py
@@ -44,6 +44,8 @@ class RedisSemanticCache(BaseCache):
         redis_url: Optional[str] = None,
         similarity_threshold: Optional[float] = None,
         embedding_model: str = "text-embedding-ada-002",
+        embedding_api_base: Optional[str] = None,
+        embedding_api_key: Optional[str] = None,
         index_name: Optional[str] = None,
         **kwargs,
     ):
@@ -86,6 +88,8 @@ class RedisSemanticCache(BaseCache):
         # While similarity: 1 = most similar, 0 = least similar
         self.distance_threshold = 1 - similarity_threshold
         self.embedding_model = embedding_model
+        self.embedding_api_base = embedding_api_base
+        self.embedding_api_key = embedding_api_key
 
         # Set up Redis connection
         if redis_url is None:
@@ -143,13 +147,18 @@ class RedisSemanticCache(BaseCache):
             List[float]: The embedding vector
         """
         # Create an embedding from prompt
+        embed_kwargs: dict = {
+            "model": self.embedding_model,
+            "input": prompt,
+            "cache": {"no-store": True, "no-cache": True},
+        }
+        if self.embedding_api_base is not None:
+            embed_kwargs["api_base"] = self.embedding_api_base
+        if self.embedding_api_key is not None:
+            embed_kwargs["api_key"] = self.embedding_api_key
         embedding_response = cast(
             EmbeddingResponse,
-            litellm.embedding(
-                model=self.embedding_model,
-                input=prompt,
-                cache={"no-store": True, "no-cache": True},
-            ),
+            litellm.embedding(**embed_kwargs),
         )
         embedding = embedding_response["data"][0]["embedding"]
         return embedding