Update app.py

app.py

@@ -1,5 +1,6 @@
 # Web Content Q&A Tool for Hugging Face Spaces
-# Features: Ingest up to 3 URLs, ask questions, get concise answers using DistilBERT with PyTorch
+# Optimized for memory constraints (2GB RAM) and 24-hour timeline
+# Features: Ingest up to 3 URLs, ask questions, get concise one-line answers using DistilBERT with PyTorch
 # Includes keyword search fallback for low-confidence QA answers
 
 import gradio as gr
@@ -57,6 +58,16 @@ model = torch.quantization.quantize_dynamic(
 # Create the QA pipeline with PyTorch
 qa_model = pipeline("question-answering", model=model, tokenizer=tokenizer, framework="pt", device=-1)  # device=-1 for CPU
 
+# Utility function to truncate text to one line
+def truncate_to_one_line(text):
+    # Split by sentence-ending punctuation and take the first sentence
+    sentences = re.split(r'[.!?]+', text.strip())
+    first_sentence = sentences[0].strip() if sentences else text.strip()
+    # If the sentence is too long, truncate to 100 characters
+    if len(first_sentence) > 100:
+        first_sentence = first_sentence[:100].rsplit(' ', 1)[0] + "..."
+    return first_sentence if first_sentence else "No answer available."
+
 # Keyword search function for fallback
 def keyword_search(question, corpus, sources_list):
     stop_words = set(["what", "is", "the", "a", "an", "in", "on", "at", "for", "with", "and", "or", "but", "not", "this", "that", "these", "those", "to", "of", "it", "by", "as", "if", "when", "where", "who", "which", "how", "why"])
@@ -85,6 +96,8 @@ def keyword_search(question, corpus, sources_list):
     if best_paragraph is None:
         return "No relevant paragraph found.", None
     
+    # Truncate the paragraph to one line
+    best_paragraph = truncate_to_one_line(best_paragraph)
     return best_paragraph, best_source
 
 def ingest_urls(urls):
@@ -140,6 +153,7 @@ def answer_question(question):
     Retrieves top 3 paragraphs to improve answer accuracy.
     If total context exceeds 512 tokens (DistilBERT's max length), it will be truncated automatically.
     If QA confidence is below 0.4, falls back to keyword search.
+    Ensures answers are one line (max 100 chars).
     """
     global corpus, embeddings, sources_list
     if not corpus or embeddings is None:
@@ -153,7 +167,7 @@ def answer_question(question):
     top_k = min(2, len(corpus))  # Get top 3 paragraphs to improve accuracy
     top_indices = np.argsort(-cos_scores)[:top_k]
     
-    # Retrieve context (top 3 paragraphs)
+    # Retrieve context (top 2 paragraphs)
     contexts = [corpus[i] for i in top_indices]
     context = " ".join(contexts)  # Concatenate with space
     sources = [sources_list[i] for i in top_indices]
@@ -165,7 +179,11 @@ def answer_question(question):
     confidence = result['score']
     
     if confidence >= 0.4:
-        # Format response with answer, confidence, and sources
+        # Truncate QA answer to one line
+        answer = truncate_to_one_line(answer)
+        # Ensure at least one line
+        if not answer:
+            answer = "No answer available."
         sources_str = "\n".join(set(sources))  # Unique sources
         return f"Answer: {answer}\nConfidence: {confidence:.2f}\nSources:\n{sources_str}"
     else: