update application file

app.py

@@ -85,7 +85,7 @@ def insert_chunks(chunks):
 def expand_query(query):
     try:
         prompt = f"""Expand the following short questions into a more detailed search query
-that includes synonyms and related HR terms:
+that includes synonyms and related HR terms, but also restate the keywords clearly:
 
 {query}
 """
@@ -99,7 +99,7 @@ that includes synonyms and related HR terms:
         print("⚠️ Query expansion failed:", e)
         return query
 
-def search_weaviate(query, k=8):
+def search_weaviate(query, k=12):
     pdf_chunks = client.collections.get("PDFChunk")
     expanded_query = expand_query(query)
     query_vec = embed(expanded_query)
@@ -107,10 +107,20 @@ def search_weaviate(query, k=8):
     result = pdf_chunks.query.hybrid(  #both lexical and semantic
             query=expanded_query,
             vector=query_vec,
-            alpha=0.6,
+            alpha=0.3,
             limit=k,
             return_properties=["text", "page"]
     )
+
+    filtered_objects = []
+    for o in result.objects:
+        distance = getattr(o.metadata, "distance", None)
+        certainty = getattr(o.metadata, "certainty", None)
+
+        # Keep results above a relevance threshold
+        if (distance is None or distance < 1.2) or (certainty and certainty >0.3):
+            filtered_objects.append(o)
+            
     return [(o.properties["text"], o.metadata.distance)for o in result.objects] 
 
 def rerank_chunks_with_llm(query, chunks):
@@ -128,7 +138,8 @@ def rerank_chunks_with_llm(query, chunks):
 
     rerank_prompt = f"""
 You are a precise HR assistant that ranks excerpts 
-from a staff handbook by how relevant they are to the user's question
+from a staff handbook by how relevant they are to the user's question.
+You must rank excerpts that directly answer the user's question higher than those that merely discuss related topics.
 
     Question: {query}
 
@@ -165,14 +176,14 @@ from a staff handbook by how relevant they are to the user's question
     return ordered_chunks
 
 def ask_question(query):
-    chunks = search_weaviate(query, k=8)
+    chunks = search_weaviate(query, k=12)
     reranked_chunks = rerank_chunks_with_llm(query, chunks)
 
     # Use top three after reranking
-    context = "\n\n---\n\n".join(reranked_chunks[:3])
+    context = "\n\n---\n\n".join(reranked_chunks[:4])
 
     prompt = f"""
-You are an HR assitant answering questions from the staff handbook.
+You are an HR assistant answering questions from the staff handbook.
 Use only the following content to answer accurately and concisely:
     {context}