Update app.py

app.py

@@ -18,7 +18,7 @@ import numpy as np
 import wikipedia
 
 # === Configuration ===
-HF_TOKEN = os.getenv("HF") 
+HF_TOKEN = os.getenv("HF")
 MODEL_NAME = "meta-llama/Llama-3.1-8B-Instruct"
 client = InferenceClient(model=MODEL_NAME, token=HF_TOKEN)
 
@@ -125,7 +125,7 @@ def fetch_text(url):
         soup = BeautifulSoup(doc.summary(), "html.parser")
         text = " ".join(p.get_text() for p in soup.find_all("p")).strip()
         return text, url
-    except:
+    except Exception as e:
         return "", url
 
 def scrape_and_save(query):
@@ -204,13 +204,18 @@ def answer_from_context(question):
 
     prompt = f"""
 Today's date is {datetime.utcnow().date()}.
-Use context and memory to answer and summarize the following question using fully finished lines, clear, and grammatically correct finished sentences. Ensure the response is factually accurate, complete, well-organized, and easy to understand. Avoid repeating information, unfinished sentences, and keep it concise but informative.
+Use context and memory to answer and summarize the following question using fullly finished lines end with., clear, and grammatically correct finish sentences. Ensure that the response is factually accurate, complete, well-organized, finish sentences and easy to understand. Avoid repeating information,unfinish sentences and keep the response concise while still being informative.
+
 [CONTEXT]
 {context}
+
 [MEMORY]
 {memory_prompt}
+
 [QUESTION]
+Answer and summarize the following question using fullly finish linesens end with., clear, and grammatically correct finish sentences. Ensure that the response is factually accurate, complete, well-organized, finish sentences and easy to understand. Avoid repeating information, unfinish sentences, and keep the response concise while still being informative.
 {question}
+
 [ANSWER]
 """
     try:
@@ -234,25 +239,29 @@ Use context and memory to answer and summarize the following question using full
 def needs_web_search_llm(question):
     prompt = f"""
 You are a helpful assistant that classifies whether a question requires a web search or external data.
+
 Question: "{question}"
+
 Answer with only "YES" if a web search is needed or "NO" if not.
 """
     try:
         response = client.text_generation(prompt, max_new_tokens=10)
         return "YES" in response.strip().upper()
-    except Exception:
+    except Exception as e:
         return False
 
 def is_general_knowledge_question(question):
     prompt = f"""
 You are a classifier. Determine if the question below can be answered using general world knowledge, like an encyclopedia or Wikipedia.
+
 Question: "{question}"
+
 Answer with "YES" if it is general knowledge. Otherwise answer "NO".
 """
     try:
         response = client.text_generation(prompt, max_new_tokens=10)
         return "YES" in response.strip().upper()
-    except Exception:
+    except Exception as e:
         return False
 
 def get_wikipedia_summary(query, sentences=3):
@@ -263,9 +272,10 @@ def get_wikipedia_summary(query, sentences=3):
         return f"Ambiguous question. Possible topics: {', '.join(e.options[:5])}"
     except wikipedia.exceptions.PageError:
         return "No Wikipedia article found for that topic."
-    except Exception:
+    except Exception as e:
         return "Error accessing Wikipedia."
 
+# === Semantic Scholar API integration ===
 def semantic_scholar_search(query, max_results=5):
     params = {
         "query": query,
@@ -277,7 +287,8 @@ def semantic_scholar_search(query, max_results=5):
         resp.raise_for_status()
         data = resp.json()
         papers = data.get("data", [])
-        texts, urls = [], []
+        texts = []
+        urls = []
         for p in papers:
             title = p.get("title", "")
             abstract = p.get("abstract", "")
@@ -299,38 +310,64 @@ def semantic_scholar_search(query, max_results=5):
         return "", []
 
 def is_research_question(question):
-    research_keywords = ["study", "paper", "evidence", "method", "experiment", "data", "results", "findings", "theory", "analysis"]
+    # Simple heuristic to detect research/scientific questions
+    keywords = [
+    "research", "study", "paper", "findings", "experiment", "scientific", "evidence", "meta-analysis",
+    "hypothesis", "literature review", "case study", "theory", "framework", "methodology", "analysis",
+    "data", "observation", "results", "variables", "survey", "questionnaire", "sampling", "experiment design",
+    "quantitative", "qualitative", "mixed methods", "statistical", "inference", "regression", "correlation",
+    "interview", "focus group", "coding", "themes", "interpretation", "reliability", "validity", "bias",
+    "significance", "conclusion", "discussion", "implications", "limitations", "future research", "peer review",
+    "publication", "citation", "replication", "protocol", "ethics", "IRB", "research question", "objective",
+    "aim", "problem statement", "gap", "contribution", "novelty", "originality", "dataset", "case", "fieldwork",
+    "observational", "experimental", "review", "systematic review", "control group", "randomized", "longitudinal",
+    "cross-sectional", "data analysis", "research design", "conceptual", "empirical", "exploratory", "descriptive",
+    "causal", "predictive", "construct", "operationalization", "dependent variable", "independent variable",
+    "mediator", "moderator", "association", "impact", "effect", "relationship", "outcome", "measure", "coding scheme"
+]
+
     q_lower = question.lower()
-    return any(kw in q_lower for kw in research_keywords)
+    return any(kw in q_lower for kw in keywords)
 
 def ask(q):
+    # Check if research/scientific question and use Semantic Scholar
     if is_research_question(q):
         context, sources = semantic_scholar_search(q)
         if context:
             answer, sources, _ = answer_from_context(q)
-            return answer, "\n".join(f"- {url}" for url in sources)
+            sources_text = "\n".join(f"- {url}" for url in sources)
+            return answer, sources_text
+        # fallback to regular web search if semantic scholar fails
         context, sources = scrape_and_save(q)
         answer, sources, _ = answer_from_context(q)
-        return answer, "\n".join(f"- {url}" for url in sources)
+        sources_text = "\n".join(f"- {url}" for url in sources)
+        return answer, sources_text
 
+    # General knowledge questions use Wikipedia
     if is_general_knowledge_question(q):
         return get_wikipedia_summary(q), "Source: Wikipedia"
 
+    # Check if we already have context stored with sufficient similarity
     _, _, avg_sim = retrieve_context_from_chunks(q)
+
+    # Check if web search is needed or context similarity too low
     intent_search = needs_web_search_llm(q)
 
     if intent_search or avg_sim < MIN_CONTEXT_SIMILARITY:
         context, sources = scrape_and_save(q)
         answer, sources, _ = answer_from_context(q)
-        return answer, "\n".join(f"- {url}" for url in sources)
+        sources_text = "\n".join(f"- {url}" for url in sources)
     else:
-        prompt = f"<|user|>\n{q.strip()}\n<|assistant|>\n"
+        # Use model to answer from prompt only
+        prompt = f"<|user|>\n Answer  and summarize the following question using fullly finish lines end with. , clear, and grammatically correct finish sentences. Ensure that the response is factually accurate, complete, well-organized, finish stances, and easy to understand. Avoid repeating information, unfinish sentences, and keep the response concise while still being informative.:\n{q.strip()}\n<|assistant|>\n"
         try:
             response = client.text_generation(prompt, max_new_tokens=512)
             answer = response.strip().split("<|assistant|>")[-1].strip()
         except Exception as e:
             answer = f"Error: {e}"
-        return answer, ""
+        sources_text = ""
+
+    return answer, sources_text
 
 # === Gradio UI ===
 with gr.Blocks() as demo:
@@ -353,4 +390,4 @@ if __name__ == '__main__':
         question = " ".join(sys.argv[1:])
         print(ask(question))
     else:
-        demo.launch()
+        demo.launch()