Filter by rank

app.py

@@ -129,7 +129,7 @@ def semantic_search(message):
 
     gradio.Info("Performing semantic search...", duration=5)
 
-    results = extractor.extract(message, n_results=25, rerank=True)
+    results = extractor.extract(message, n_results=5, rerank=True)
 
     print(results)
 
@@ -196,7 +196,7 @@ with gradio.Blocks() as demo:
 
     # Options
     model_selector = gradio.Dropdown([(item["name"], item["repo"]) for item in model_options], value="daniel-dona/sparql-model-era-lora-128-qwen3-4b", render=False, interactive=True, label="Model", info="Base model provided as reference, SFT model is trained on generated datasets, GRPO model is reinforced on to of SFT.")
-    model_temperature = gradio.Slider(0, 1, render=False, step=0.1, value=0.6, label="Temperature", info="Ajust model variability, with a value of 0, the model use greedy decoding.")
+    model_temperature = gradio.Slider(0, 1, render=False, step=0.1, value=0, label="Temperature", info="Ajust model variability, with a value of 0, the model use greedy decoding.")
     sparql_endpoint = gradio.Textbox(value=SPARQL_ENDPOINT, render=False, interactive=True, label="SPARQL endpoint", info="SPARQL endpoint to send the generate queries to fetch results.")
     model_semantic = gradio.Checkbox(value=True, render=False, interactive=True, label="Enable semantic entity lookup", info="Use embeddings and reranking model to retrieve relevant objects.")
     model_thinking = gradio.Checkbox(value=False, render=False, interactive=True, label="Enable thinking", info="Use thinking mode in the Jinja chat template, mostly for GRPO experiments.")

src/SemanticSearch.py

@@ -201,7 +201,7 @@ class SemanticSearch:
         self.collection = client.get_collection(name=self.collection_name)
 
     
-    def extract(self, nlq: str, n_results:int=10, n_candidates:int=50, str_match:bool=True, rerank:bool=True):
+    def extract(self, nlq: str, n_results:int=10, n_candidates:int=50, rerank:bool=True, rank_cut:float=0.0):
 
         embedding = self.get_text_embeddings_local([nlq])[0].tolist()
 
@@ -229,6 +229,8 @@ class SemanticSearch:
                 {"rank": result[0], "document": result[1], "uri": result[2]} 
                 for result in zip(results["rank"][0], results["documents"][0], results["uris"][0])
             ], key=lambda x: x["rank"], reverse=True)
+            
+            results = [result for result in results if result["rank"] >= rank_cut]
 
         else: