📐 feat: add EMBEDDINGS_DIMENSIONS env for text-embedding-3 models

README.md

@@ -106,6 +106,7 @@ The following environment variables are required to run the application:
     - bedrock: "amazon.titan-embed-text-v1"
     - google_genai: "gemini-embedding-001"
 - `EMBEDDINGS_CHUNK_SIZE`: (Optional) The chunk size used by the OpenAI and Azure embeddings clients to limit the number of inputs per request. Default value is `200`.
+- `EMBEDDINGS_DIMENSIONS`: (Optional) Output vector size to request from the embedding model. Only honored by the `openai` and `azure` providers, and only supported by `text-embedding-3-*` models. Leave unset to use the model's native dimensionality (1536 for `text-embedding-3-small`, 3072 for `text-embedding-3-large`). Setting a smaller value (e.g. `512`, `1024`) trades some retrieval quality for lower storage cost and faster similarity search. Note: do not change this on an existing collection — all vectors in a `pgvector` column must share the same dimensionality.
 - `RAG_AZURE_OPENAI_API_VERSION`: (Optional) Default is `2023-05-15`. The version of the Azure OpenAI API.
 - `RAG_AZURE_OPENAI_API_KEY`: (Optional) The API key for Azure OpenAI service.
     - Note: `AZURE_OPENAI_API_KEY` will work but `RAG_AZURE_OPENAI_API_KEY` will override it in order to not conflict with LibreChat setting.

app/config.py

@@ -218,29 +218,35 @@ async def dispatch(self, request, call_next):
 ## Embeddings
 
 
-def init_embeddings(provider, model):
+def init_embeddings(provider, model, dimensions=None):
     if provider == EmbeddingsProvider.OPENAI:
         from langchain_openai import OpenAIEmbeddings
 
-        return OpenAIEmbeddings(
+        kwargs = dict(
             model=model,
             api_key=RAG_OPENAI_API_KEY,
             openai_api_base=RAG_OPENAI_BASEURL,
             openai_proxy=RAG_OPENAI_PROXY,
             chunk_size=EMBEDDINGS_CHUNK_SIZE,
             check_embedding_ctx_length=RAG_CHECK_EMBEDDING_CTX_LENGTH,
         )
+        if dimensions is not None:
+            kwargs["dimensions"] = dimensions
+        return OpenAIEmbeddings(**kwargs)
     elif provider == EmbeddingsProvider.AZURE:
         from langchain_openai import AzureOpenAIEmbeddings
 
-        return AzureOpenAIEmbeddings(
+        kwargs = dict(
             azure_deployment=model,
             api_key=RAG_AZURE_OPENAI_API_KEY,
             azure_endpoint=RAG_AZURE_OPENAI_ENDPOINT,
             api_version=RAG_AZURE_OPENAI_API_VERSION,
             chunk_size=EMBEDDINGS_CHUNK_SIZE,
             check_embedding_ctx_length=RAG_CHECK_EMBEDDING_CTX_LENGTH,
         )
+        if dimensions is not None:
+            kwargs["dimensions"] = dimensions
+        return AzureOpenAIEmbeddings(**kwargs)
     elif provider == EmbeddingsProvider.HUGGINGFACE:
         from langchain_huggingface import HuggingFaceEmbeddings
 
@@ -298,6 +304,18 @@ def init_embeddings(provider, model):
     get_env_variable("EMBEDDINGS_PROVIDER", EmbeddingsProvider.OPENAI.value).lower()
 )
 
+# Only parse EMBEDDINGS_DIMENSIONS for providers that honor it (OpenAI / Azure).
+# Parsing unconditionally at import would turn an unrelated stale env var
+# (e.g. EMBEDDINGS_DIMENSIONS=foo left over from an OpenAI deployment) into a
+# hard boot failure under bedrock / hf / ollama / etc., even though those
+# providers silently ignore the value.
+EMBEDDINGS_DIMENSIONS = None
+
+if EMBEDDINGS_PROVIDER in (EmbeddingsProvider.OPENAI, EmbeddingsProvider.AZURE):
+    _embeddings_dimensions_raw = get_env_variable("EMBEDDINGS_DIMENSIONS", None)
+    if _embeddings_dimensions_raw not in (None, ""):
+        EMBEDDINGS_DIMENSIONS = int(_embeddings_dimensions_raw)
+
 if EMBEDDINGS_PROVIDER == EmbeddingsProvider.OPENAI:
     EMBEDDINGS_MODEL = get_env_variable("EMBEDDINGS_MODEL", "text-embedding-3-small")
     # 1000 is the default chunk size for OpenAI, but this causes API rate limits to be hit
@@ -328,7 +346,9 @@ def init_embeddings(provider, model):
 else:
     raise ValueError(f"Unsupported embeddings provider: {EMBEDDINGS_PROVIDER}")
 
-embeddings = init_embeddings(EMBEDDINGS_PROVIDER, EMBEDDINGS_MODEL)
+embeddings = init_embeddings(
+    EMBEDDINGS_PROVIDER, EMBEDDINGS_MODEL, dimensions=EMBEDDINGS_DIMENSIONS
+)
 
 logger.info(f"Initialized embeddings of type: {type(embeddings)}")