fix: enforce 'NONE' responses to stop conversational pollution in invalid keywords

app.py

@@ -9,59 +9,42 @@ from langchain_core.prompts import ChatPromptTemplate
 from langchain_openai import ChatOpenAI
 
 # ==========================================================
-# 1. LOAD DATA & OPTIMIZED INDEX GENERATOR (REDUCES TOKENS)
+# 1. LOAD DATA & OPTIMIZED INDEX GENERATOR
 # ==========================================================
 with open("gcmd_hierarchy.json", "r") as f:
     gcmd_data = json.load(f)
 
 def build_gcmd_indices(gcmd_json):
-    """
-    Parses the GCMD hierarchy exactly like the notebook but strips 
-    heavy definitions to prevent 429 Token Rate Limit errors.
-    """
     topic_list = []
     sub_tree_indices = {}
-
-    # The root node level is "Category" (EARTH SCIENCE)
     topics = gcmd_json.get("children", [])
     
     for topic_node in topics:
         topic_name = topic_node.get("name", "").upper()
         topic_list.append(topic_name)
-        
-        # Collect paths inside this specific topic
         collected_paths = []
         
         def recurse_sub_tree(node, current_path=""):
             node_name = node.get("name", "")
-            # Append current node name to path
             node_path = f"{current_path} > {node_name}" if current_path else node_name
-            
-            # If it's a Variable or leaf term, save the path only
             if "Variable" in node.get("level", "") or not node.get("children"):
                 collected_paths.append(node_path)
-            
-            # Dig deeper into children
             for child in node.get("children", []):
                 recurse_sub_tree(child, node_path)
 
-        # Start the recursion from the children of this Topic (the 'Terms')
         for term_node in topic_node.get("children", []):
             recurse_sub_tree(term_node, current_path="")
             
-        # Format the collected paths into a clean lookup layout for the prompt
         sub_tree_indices[topic_name] = "\n".join([f"- {path}" for path in collected_paths])
-        
     return topic_list, sub_tree_indices
 
 VALID_TOPICS, SUB_TREE_LOOKUP = build_gcmd_indices(gcmd_data)
 
 # ==========================================================
-# 2. VERBATIM WORKFLOW FROM YOUR NOTEBOOK
+# 2. UPDATED LANGGRAPH WORKFLOW WITH NULL-ENFORCEMENT
 # ==========================================================
 
 def merge_lists(left: list, right: list) -> list:
-    """A reducer function that merges list contents across parallel branches."""
     return list(set((left or []) + (right or [])))
 
 class MultiTopicState(TypedDict):
@@ -75,13 +58,17 @@ class TopicsChoice(BaseModel):
     topics: List[str] = Field(description="List of matching topic areas from the allowed dataset.")
 
 def route_multi_topic(state: MultiTopicState):
-    """Step 1: Identify ALL relevant high-level topics for the paper."""
+    """Step 1: Identify strictly relevant high-level topics (Tightened to prevent over-routing)."""
     llm = ChatOpenAI(model="gpt-4o", temperature=0)
     structured_llm = llm.with_structured_output(TopicsChoice)
     
     prompt = ChatPromptTemplate.from_messages([
-        ("system", f"You are an expert science cataloger. Identify ALL relevant major topic areas that apply to this paper. Choose ONLY from: {', '.join(VALID_TOPICS)}"),
-        ("user", "Title: {title}\nAbstract: {abstract}\n\nSelect all relevant Topics as a structured list.")
+        ("system", (
+            f"You are an expert science cataloger. Identify only the core major topic areas that "
+            f"DIRECTLY and primary apply to this paper. Do NOT select topics that are only tangentially "
+            f"referenced or inferred. Choose ONLY from: {', '.join(VALID_TOPICS)}"
+        )),
+        ("user", "Title: {title}\nAbstract: {abstract}\n\nSelect the highly relevant Topics as a structured list.")
     ])
     
     result = structured_llm.invoke(prompt.format(title=state["title"], abstract=state["abstract"]))
@@ -93,21 +80,30 @@ def route_multi_topic(state: MultiTopicState):
     return {"chosen_topics": valid_selected}
 
 def classify_individual_topic(topic_name: str):
-    """A dynamic factory function that returns a custom node runner for a specific topic."""
+    """A dynamic factory function with strict enforcement against conversational prose."""
     
     def node_runner(state: MultiTopicState):
         llm = ChatOpenAI(model="gpt-4o", temperature=0)
         target_sub_tree = SUB_TREE_LOOKUP.get(topic_name, "")
         
         prompt = ChatPromptTemplate.from_messages([
-            ("system", f"You are a specialist in {topic_name} data mapping. Extract exact keyword pathways present in the provided list."),
+            ("system", (
+                f"You are a specialist in {topic_name} data mapping. Extract exact keyword pathways present in the provided list.\n"
+                f"CRITICAL RULES:\n"
+                f"1. If absolutely no keywords from the valid path list match this paper, reply with exactly the word 'NONE'.\n"
+                f"2. Do NOT write sentences, do NOT explain your reasoning, and do NOT say 'there are no matching entries'. Your output must only be a comma-separated list of keywords, or the single token 'NONE'."
+            )),
             ("user", "Title: {title}\nAbstract: {abstract}\n\nValid Paths:\n{sub_tree}\n\nReturn exact matching entries as a comma-separated list.")
         ])
         
         response = llm.invoke(prompt.format(title=state["title"], abstract=state["abstract"], sub_tree=target_sub_tree))
-        raw_keywords = [k.strip() for k in response.content.split(",") if k.strip()]
         
-        # Immediate validation pass inside the node branch
+        # Split and clean tokens, explicitly ignoring any 'NONE' or empty responses
+        raw_keywords = [
+            k.strip() for k in response.content.split(",") 
+            if k.strip() and k.strip().upper() != "NONE"
+        ]
+        
         valid_set = set()
         for line in target_sub_tree.split("\n"):
             if line.strip():
@@ -122,7 +118,6 @@ def classify_individual_topic(topic_name: str):
     return node_runner
 
 def parallel_router(state: MultiTopicState) -> List[str]:
-    """Tells LangGraph to trigger multiple sub-nodes simultaneously based on state."""
     return [f"classify_{topic.lower()}" for topic in state["chosen_topics"]]
 
 # Assemble the Graph
@@ -149,15 +144,13 @@ workflow.add_edge("classify_fallback", END)
 app = workflow.compile()
 
 # ==========================================================
-# 3. GRADIO USER INTERFACE (WITH PRESENTATION FORMATTING)
+# 3. GRADIO USER INTERFACE
 # ==========================================================
 def run_agent_classifier(title, abstract):
     if not title or not abstract:
         return "Please fill out both Title and Abstract fields.", "N/A", "N/A"
         
     inputs = {"title": title, "abstract": abstract}
-    
-    # Execute graph synchronously
     output = app.invoke(inputs)
     
     chosen_topics = output.get("chosen_topics", [])
@@ -166,7 +159,6 @@ def run_agent_classifier(title, abstract):
     
     topics_str = ", ".join(chosen_topics)
     
-    # Format output keywords to prepend their corresponding parent Topic
     formatted_keywords = []
     for kw in predicted_kws:
         matched_topic = "UNKNOWN"
@@ -182,7 +174,6 @@ def run_agent_classifier(title, abstract):
     else:
         keywords_str = "\n".join(formatted_keywords)
         
-    # Format separate display for caught and filtered hallucinations
     if not invalid_kws:
         invalid_str = "None! The agent validation pass achieved 100% data integrity."
     else: