Update app.py

app.py

@@ -9,7 +9,7 @@ import faiss
 import torch
 
 # ===============================
-# EMBEDDING MODEL
+# EMBEDDING MODEL SETUP
 # ===============================
 model_name = "sentence-transformers/all-MiniLM-L6-v2"
 tokenizer = AutoTokenizer.from_pretrained(model_name)
@@ -22,16 +22,14 @@ def get_embeddings(texts):
     with torch.no_grad():
         outputs = embedding_model(**inputs)
     embeddings = outputs.last_hidden_state[:, 0].cpu().numpy()
+    # Normalize embeddings to unit length for cosine similarity
+    embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)
     return embeddings
 
-def normalize_embeddings(embeddings):
-    norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
-    return embeddings / (norms + 1e-10)
-
 # ===============================
 # TEXT CHUNKING
 # ===============================
-def chunk_text(text, chunk_size=400, overlap=100):
+def chunk_text(text, chunk_size=500, overlap=50):
     chunks = []
     start = 0
     while start < len(text):
@@ -46,7 +44,7 @@ def chunk_text(text, chunk_size=400, overlap=100):
 index_path = "faiss_index.pkl"
 document_texts_path = "document_texts.pkl"
 document_texts = []
-embedding_dim = 384
+embedding_dim = 384  # For all-MiniLM-L6-v2
 
 if os.path.exists(index_path) and os.path.exists(document_texts_path):
     try:
@@ -78,7 +76,6 @@ def extract_text_from_docx(path):
     try:
         doc = Document(path)
         text = "\n".join([para.text for para in doc.paragraphs])
-        print(f"Extracted DOCX text preview: {text[:500]}")  # Preview first 500 chars for debug
     except Exception as e:
         print(f"DOCX error: {e}")
     return text
@@ -97,7 +94,6 @@ def upload_document(file):
 
     chunks = chunk_text(text)
     chunk_embeddings = get_embeddings(chunks)
-    chunk_embeddings = normalize_embeddings(chunk_embeddings)
     index.add(np.array(chunk_embeddings).astype('float32'))
     document_texts.extend(chunks)
 
@@ -113,30 +109,29 @@ def upload_document(file):
 # ===============================
 qa_pipeline = pipeline("text2text-generation", model="google/flan-t5-base")
 
-def generate_answer_from_file(query, top_k=5):
+def generate_answer_from_file(query, top_k=7):
     if not document_texts:
         return "No documents indexed yet."
 
     query_vector = get_embeddings(query).astype("float32")
-    query_vector = normalize_embeddings(query_vector)
     scores, indices = index.search(query_vector, k=top_k)
     retrieved_chunks = [document_texts[i] for i in indices[0]]
     context = "\n\n".join(retrieved_chunks)
 
     prompt = (
-        "You are a helpful assistant reading student notes or textbook passages.\n\n"
-        "Based on the context provided, answer the question accurately.\n\n"
+        "You are a helpful and precise assistant reading student notes or textbook passages.\n\n"
+        "Based on the context provided, answer the question accurately and in detail using full sentences.\n\n"
         "### Example\n"
         "Context:\nArtificial systems are created by people. These systems are designed to perform specific tasks, improve efficiency, and solve problems. Examples include knowledge systems, engineering systems, and social systems.\n\n"
         "Question: What is an Artificial System?\n"
-        "Answer: Artificial systems are systems created by humans to perform specific tasks, improve efficiency, and solve problems. They include systems like knowledge systems, engineering systems, and social systems.\n\n"
+        "Answer: Artificial systems are systems created by humans to perform specific tasks, improve efficiency, and solve problems. They include systems such as knowledge systems, engineering systems, and social systems.\n\n"
         "### Now answer this\n"
         f"Context:\n{context}\n\n"
         f"Question: {query}\n"
-        f"Answer:"
+        "Answer:\nPlease answer ONLY based on the context above without adding extra information."
     )
 
-    result = qa_pipeline(prompt, max_length=512, do_sample=False)[0]['generated_text']
+    result = qa_pipeline(prompt, max_length=700, do_sample=False)[0]['generated_text']
     return result.strip()
 
 # ===============================
@@ -160,3 +155,4 @@ search_interface = gr.Interface(
 
 app = gr.TabbedInterface([upload_interface, search_interface], ["Upload", "Ask"])
 app.launch()
+