update application file to include reranker and change embedding model

app.py

@@ -7,6 +7,7 @@ from pypdf import PdfReader
 from pathlib import Path
 from weaviate.auth import AuthApiKey
 from dotenv import load_dotenv
+import re
 
 #Setup
 OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
@@ -67,7 +68,7 @@ def setup_schema():
 #Create embeddings and Store in Vector DB
 def embed(text):
     return openai_client.embeddings.create(
-        model = "text-embedding-3-small",
+        model = "text-embedding-3-large",
         input=text
     ).data[0].embedding
 
@@ -81,7 +82,7 @@ def insert_chunks(chunks):
         )
 
 # Querying
-def search_weaviate(query, k=3):
+def search_weaviate(query, k=5):
     pdf_chunks = client.collections.get("PDFChunk")
     query_vec = embed(query)
     
@@ -90,22 +91,83 @@ def search_weaviate(query, k=3):
             limit=k,
             return_properties=["text", "page"]
     )
-    return [o.properties["text"]for o in result.objects] 
+    return [(o.properties["text"], o.metadata.distance)for o in result.objects] 
+
+def rerank_chunks_with_llm(query, chunks):
+    """
+    Rerank retrieved chunks using GPT reasoning.
+    Returns a list of chunks ordered in descending order
+    """
+
+    #Build a short reranking prompt
+    chunk_list = "\n\n".join([f"[{i+1}] {text[:400].strip().replace('\n', ' ')}..."
+                            for i, (text, _) in enumerate(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
+
+    Question: {query}
+
+    Excerpts:
+    {chunk_list}
+
+    Return only the list of excerpt numbers, separated by commas, in descending order of relevance.
+    Example: 3, 1, 2
+    """
+
+    #Run LLM model
+    response = openai_client.chat.completions.create(
+        model="gpt-4.1-mini",
+        messages=[
+            {"role": "system", "content": "You are a factual and consistent reranker."},
+            {"role": "user", "content": rerank_prompt}
+        ],
+        temperature = 0
+    )
+    
+    text_output = response.choices[0].message.content.strip()
+    print(f"🔎 Reranker raw output: {text_output}")  # optional
+
+    # extract numbers safely
+    order = [int(x) for x in re.findall(r'\d+', text_output )]
+    order = [i for i in order if 1 <= i <= len(chunks)] #ensure valid range
+
+    # fallback: if model fails to output indices, return original order
+    if not order:
+        order = list(range(1, len(chunks) + 1))
+    
+    # Return reordered text chunks
+    ordered_chunks = [chunks[i-1][0] for i in order]
+    return ordered_chunks
 
 def ask_question(query):
-    chunks = search_weaviate(query)
-    context = "\n".join(chunks)
+    chunks = search_weaviate(query, k=5)
+    reranked_chunks = rerank_chunks_with_llm(query, chunks)
+
+    # Use top three after reranking
+    context = "\n\n---\n\n".join(reranked_chunks[:3])
 
     prompt = f"""
-    Answer the question based only on the following PDF content:\n\n{context}\n\n
-    Question: {query}\nAnswer:
+    You are an HR assitant answering questions from the staff handbook.
+Use only the following content to answer accurately and concisely:
+    {context}
+
+    Question: {query}
+    Answer:
     """
 
     response = openai_client.chat.completions.create(
         model="gpt-4.1-mini",
-        messages=[{"role": "user", "content": prompt}]
+        messages=[
+            {"role": "system", "content": "You are a helpful HR assistant."},
+            {"role": "user", "content": prompt}
+    ],
+    temperature=0
     )
-    return response.choices[0].message.content #explain choices
+
+    return response.choices[0].message.content.strip() 
 
 #Gradio App
 def process_pdf(pdf_file):