From 04d558e75f660eaadd506a25e8c2c952f95b2ff5 Mon Sep 17 00:00:00 2001 From: Emerson Gomes Date: Fri, 6 Dec 2024 18:47:11 -0600 Subject: [PATCH] Correct Vertex Embedding Model Data/Prices (#7069) * Correct embedding model prices * Corrected max_tokens * Added text-embedding-005 --- model_prices_and_context_window.json | 36 +++++++++++++++++++++------- 1 file changed, 27 insertions(+), 9 deletions(-) diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 481b1fd229..487165cad6 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -3215,10 +3215,22 @@ "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing" }, "text-embedding-004": { - "max_tokens": 3072, - "max_input_tokens": 3072, + "max_tokens": 2048, + "max_input_tokens": 2048, "output_vector_size": 768, - "input_cost_per_token": 0.00000000625, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, + "output_cost_per_token": 0, + "litellm_provider": "vertex_ai-embedding-models", + "mode": "embedding", + "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models" + }, + "text-embedding-005": { + "max_tokens": 2048, + "max_input_tokens": 2048, + "output_vector_size": 768, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, "output_cost_per_token": 0, "litellm_provider": "vertex_ai-embedding-models", "mode": "embedding", @@ -3228,7 +3240,8 @@ "max_tokens": 2048, "max_input_tokens": 2048, "output_vector_size": 768, - "input_cost_per_token": 0.00000000625, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, "output_cost_per_token": 0, "litellm_provider": "vertex_ai-embedding-models", "mode": "embedding", @@ -3238,7 +3251,8 @@ "max_tokens": 3072, "max_input_tokens": 3072, "output_vector_size": 768, - "input_cost_per_token": 0.00000000625, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, "output_cost_per_token": 0, "litellm_provider": "vertex_ai-embedding-models", "mode": "embedding", @@ -3248,7 +3262,8 @@ "max_tokens": 3072, "max_input_tokens": 3072, "output_vector_size": 768, - "input_cost_per_token": 0.00000000625, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, "output_cost_per_token": 0, "litellm_provider": "vertex_ai-embedding-models", "mode": "embedding", @@ -3258,7 +3273,8 @@ "max_tokens": 3072, "max_input_tokens": 3072, "output_vector_size": 768, - "input_cost_per_token": 0.00000000625, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, "output_cost_per_token": 0, "litellm_provider": "vertex_ai-embedding-models", "mode": "embedding", @@ -3268,7 +3284,8 @@ "max_tokens": 3072, "max_input_tokens": 3072, "output_vector_size": 768, - "input_cost_per_token": 0.00000000625, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, "output_cost_per_token": 0, "litellm_provider": "vertex_ai-embedding-models", "mode": "embedding", @@ -3278,7 +3295,8 @@ "max_tokens": 3072, "max_input_tokens": 3072, "output_vector_size": 768, - "input_cost_per_token": 0.00000000625, + "input_cost_per_character": 0.000000025, + "input_cost_per_token": 0.0000001, "output_cost_per_token": 0, "litellm_provider": "vertex_ai-embedding-models", "mode": "embedding",