[WIP] Fix AL workflow routing logic

RAGManager/app/agents/nodes/agent_host.py

@@ -1,12 +1,8 @@
 """Nodo 1: Agent Host - Entry point that saves initial context."""
 
 import logging
-from uuid import UUID, uuid4
 
 from app.agents.state import AgentState
-from app.core.config import settings
-from app.core.database_connection import SessionLocal
-from app.models.chat import ChatMessage, ChatSession
 
 logger = logging.getLogger(__name__)
 
@@ -29,7 +25,7 @@ def agent_host(state: AgentState) -> AgentState:
         Updated state with chat_session_id, chat_messages, and initial_context set
     """
     updated_state = state.copy()
-    initial_message = state["messages"][-1]
+    initial_message = state["messages"][-1] if state["messages"] else None
     updated_state["initial_context"] = (
         initial_message.content if initial_message else ""
     )

RAGManager/app/agents/nodes/context_builder.py

@@ -1,10 +1,19 @@
 """Nodo 6: Context Builder - Enriches query with retrieved context."""
 
-from app.agents.state import AgentState
 from langchain_core.messages import SystemMessage
 from langchain_openai import ChatOpenAI
 
-llm = ChatOpenAI(model="gpt-5-nano")
+from app.agents.state import AgentState
+
+_llm: ChatOpenAI | None = None
+
+
+def _get_llm() -> ChatOpenAI:
+    """Lazy initialization of LLM instance."""
+    global _llm
+    if _llm is None:
+        _llm = ChatOpenAI(model="gpt-5-nano")
+    return _llm
 
 
 def context_builder(state: AgentState) -> dict:
@@ -21,7 +30,7 @@ def context_builder(state: AgentState) -> dict:
         state: Agent state containing paraphrased_text and relevant_chunks
 
     Returns:
-        Updated state with enriched_query and primary_response set
+        dict: A dictionary with a "messages" key containing the LLM response
     """
     # TODO: Implement context building and primary LLM call
     # This should:
@@ -31,13 +40,11 @@ def context_builder(state: AgentState) -> dict:
     # 4. Store the LLM response in primary_response
 
     # Placeholder: For now, we'll create a simple enriched query
-    updated_state = state.copy()
-    paraphrased = state.get("paraphrased_text", "")
     chunks = state.get("relevant_chunks", [])
 
     # Build enriched query with context
     context_section = "\n\n".join(chunks) if chunks else "No relevant context found."
-    
+
     system_content = f"""You are a helpful assistant. Use the following context to answer the user's question.
 If the answer is not in the context, say you don't know.
 
@@ -47,6 +54,6 @@ def context_builder(state: AgentState) -> dict:
     messages = [SystemMessage(content=system_content)] + state["messages"]
 
     # Call Primary LLM
-    response = llm.invoke(messages)
+    response = _get_llm().invoke(messages)
 
     return {"messages": [response]}

RAGManager/app/agents/nodes/fallback.py

@@ -1,34 +1,42 @@
-"""Nodo 3: Fallback Inicial - Initial fallback processing."""
+"""Nodo 3: Fallback - Fallback processing for malicious prompts and risky content."""
 
 import logging
 
-from app.agents.state import AgentState
 from langchain_core.messages import SystemMessage
 from langchain_openai import ChatOpenAI
 
+from app.agents.state import AgentState
+
 logger = logging.getLogger(__name__)
 
-llm = ChatOpenAI(
-    model="gpt-5-nano",
-)
+_llm: ChatOpenAI | None = None
+
+
+def _get_llm() -> ChatOpenAI:
+    """Lazy initialization of LLM instance."""
+    global _llm
+    if _llm is None:
+        _llm = ChatOpenAI(
+            model="gpt-5-nano",
+        )
+    return _llm
 
 
-# TO DO: implementar clase nodo fallback y inicializar el llm en el init
-def fallback(state: AgentState) -> AgentState:
+def fallback(state: AgentState) -> dict:
     """
     Fallback node - Performs fallback processing.
 
     This node:
     1. Alerts about malicious prompt or PII detection
-    2. Generates an error_message from llm to show the user
+    2. Generates an error message from llm to show the user
 
     Args:
         state: Agent state containing the prompt or initial context
 
     Returns:
-        error_message
+        dict: A dictionary with a "messages" key containing a list with the generated error message from the LLM.
     """
-    
+
     # Check for PII/Risky content (from guard_final)
     if state.get("is_risky"):
         logger.warning(
@@ -38,17 +46,17 @@ def fallback(state: AgentState) -> AgentState:
             "Your job is to generate an error message in user's language explaining "
             "that the response cannot be provided because it contains sensitive or private information."
         )
-        
+
     # Check for Malicious prompt (from guard_inicial)
     elif state.get("is_malicious"):
         logger.warning(
-            "Defensive check triggered: Malicious prompt detected" 
+            "Defensive check triggered: Malicious prompt detected"
         )
         system_message_content = (
             "Your job is to generate an error message in user's language for the user "
             "explaining the database doesn't have the information to answer the user's question"
         )
-        
+
     # Generic Fallback (neither risky nor malicious)
     else:
         logger.info(
@@ -58,11 +66,11 @@ def fallback(state: AgentState) -> AgentState:
             "Your job is to generate an error message in user's language for the user "
             "explaining the database doesn't have the information to answer the user's question"
         )
-    
+
     messages = [
         SystemMessage(content=system_message_content)
     ] + state["messages"]
-    
-    error_message = llm.invoke(messages)
+
+    error_message = _get_llm().invoke(messages)
     return {"messages": [error_message]}
 

RAGManager/app/agents/nodes/guard_inicial.py

@@ -43,8 +43,12 @@ def guard_inicial(state: AgentState) -> AgentState:
         Updated state with is_malicious and error_message set
     """
     updated_state = state.copy()
-    last_message = state["messages"][-1]
-    prompt = last_message.content if last_message else ""
+    if state["messages"]:
+        last_message = state["messages"][-1]
+        prompt = last_message.content if last_message else ""
+    else:
+        last_message = None
+        prompt = ""
 
     if not prompt:
         # Empty prompt is considered safe
@@ -77,7 +81,7 @@ def guard_inicial(state: AgentState) -> AgentState:
     except Exception as e:
         # Log error details for monitoring
         logger.error("Error during jailbreak detection", exc_info=True)
-        
+
         # Check if fail-closed mode is enabled
         if settings.guardrails_fail_closed:
             # Fail-closed: treat errors as malicious to prevent bypassing detection

RAGManager/app/agents/nodes/parafraseo.py

@@ -1,10 +1,19 @@
 """Nodo 4: Parafraseo - Paraphrases user input."""
 
-from app.agents.state import AgentState
 from langchain_core.messages import SystemMessage
 from langchain_openai import ChatOpenAI
 
-llm = ChatOpenAI(model="gpt-5-nano")
+from app.agents.state import AgentState
+
+_llm: ChatOpenAI | None = None
+
+
+def _get_llm() -> ChatOpenAI:
+    """Lazy initialization of LLM instance."""
+    global _llm
+    if _llm is None:
+        _llm = ChatOpenAI(model="gpt-5-nano")
+    return _llm
 
 
 def parafraseo(state: AgentState) -> AgentState:
@@ -29,14 +38,14 @@ def parafraseo(state: AgentState) -> AgentState:
     # 3. Set paraphrased_text with the result
 
     # Paraphrase the last message using history
-    
+
     system_instruction = """You are an expert at paraphrasing user questions to be standalone and clear, given the conversation history.
 Reformulate the last user message to be a self-contained query that includes necessary context from previous messages.
 Do not answer the question, just rewrite it."""
 
     messages = [SystemMessage(content=system_instruction)] + state["messages"]
-    
-    response = llm.invoke(messages)
+
+    response = _get_llm().invoke(messages)
     updated_state = state.copy()  # Create a copy of the state to update
     updated_state["paraphrased_text"] = response.content