chore: update inference version

inference/config.py

@@ -52,7 +52,7 @@ class Config:
 
     def __init__(self):
         # version
-        self.VERSION = "v0.2.19"
+        self.VERSION = "v0.2.20"
 
         # mode
         self.MODE = load_str_env("MODE", required=True)

inference/test/test_rerank.py

@@ -79,10 +79,7 @@ async def test_rerank(self, test_data):
         for result in results:
             assert result.get("document").get("text") in self.documents
             assert result.get("relevance_score") >= 0.0
-            if "jina-colbert-v1-en" in model_schema_id:
-                assert result.get("relevance_score") <= 10.0
-            else:
-                assert result.get("relevance_score") <= 1.0
+            assert result.get("relevance_score") <= 1.0
             assert result.get("index") == self.documents.index(result.get("document").get("text"))
 
     @pytest.mark.test_id("inference_032")

inference/test/utils/utils.py

@@ -31,6 +31,8 @@
     "mistralai/mistral-small-latest",
     "groq/llama-3.1-8b-instant",
     "groq/gemma-7b",
+    "reka/reka-edge",  # This model temporarily encountered an internal server error in inference version 0.2.19
+    "jina/jina-colbert-v1-en",  # The relevance_score returned by this model lacks a defined range
 ]
 
 load_dotenv()

inference/test/utils/wildcard_test_cases.yml

@@ -33,7 +33,7 @@ wildcard_test_cases:
   - provider_id: "openrouter"
     cases:
       - model_schema_id: "openrouter/wildcard"
-        provider_model_id: "nousresearch/nous-capybara-7b:free"
+        provider_model_id: "mattshumer/reflection-70b:free"
         model_type: "chat_completion"
         streaming: True
         function_call: True