push

app/gemini_client.py

@@ -78,7 +78,7 @@ def gemini_generate_text(
     system_instruction: str,
     user_text: str,
     temperature: float = 0.0,
-    max_output_tokens: int = 4096,  # 1. Provide plenty of output space
+    max_output_tokens: int = 4096,  
     response_schema: Any | None = None,
 ) -> str:
     from google.genai import types  # type: ignore[import-untyped]
@@ -93,7 +93,7 @@ def gemini_generate_text(
             max_output_tokens=max_output_tokens,
             response_mime_type="application/json",
             response_schema=response_schema,
-            # 2. Allocate thinking depth specifically for high-throughput structures
+            
             thinking_config=types.ThinkingConfig(
                 thinking_level="minimal"  # Options: 'minimal' or 'low' to free up response tokens
             )

app/main.py

@@ -147,7 +147,7 @@ class RecommendationRequest(BaseModel):
                     "city": None,
                     "state": None,
                     "chat_history": [],
-                    "top_k_retrieval": 6,
+                    "top_k_retrieval": 30,
                     "top_n_final": 5,
                 }
             ]

app/recommendation_pipeline.py

@@ -15,9 +15,7 @@ from app.shared_models import embedding_model_name_task_b, get_embedder
 
 logger = logging.getLogger(__name__)
 
-# =========================================================
-# COMPETITION-READY STRUCTURED SCHEMA
-# =========================================================
+
 class RecommendationItem(BaseModel):
     business_id: str = Field(description="The unique identifier string of the recommended business.")
     rationale: str = Field(
@@ -34,9 +32,7 @@ class AgentRecommendationOutput(BaseModel):
     )
 
 
-# =========================================================
-# UTILS & VECTORRETRIEVAL (Maintained for Baseline Retrieval)
-# =========================================================
+
 def _resolve_catalog_path(raw: str) -> Path:
     p = Path(raw)
     if p.is_absolute():
@@ -186,7 +182,7 @@ Candidate Data Pools:
         ]
 
     results = []
-    # Build final schema objects up to your pipeline boundary parameters
+
     for i, item in enumerate(recommendation_items[:top_n]):
         results.append({
             "business_id": item.get("business_id"),
@@ -224,7 +220,7 @@ class RecommendationService:
         city: str | None = None,
         state: str | None = None,
         chat_history: list[dict[str, str]] | None = None,
-        top_k_retrieval: int = 6,
+        top_k_retrieval: int = 30,
         top_n_final: int = 5,
     ) -> dict[str, Any]:
 
@@ -234,10 +230,10 @@ class RecommendationService:
         qtext = build_query_text(persona, chat_history)
         query_emb = embed_query_text(qtext)
 
-        # Execution Step 1: Initial Contextual Candidate Retrieval (Vector Base)
+      
         candidates = self.index.retrieve(query_emb, top_k_retrieval, city, state)
 
-        # Execution Step 2: Agentic Personalization Reranking via Gemini
+       
         ranked = chat_rank_gemini(
             persona=persona,
             chat_history=chat_history,