Update xxxx

my_tools.py

@@ -18,6 +18,11 @@ from llama_index.core.tools import FunctionTool
 from llama_index.core.agent import ReActAgent
 from llama_index.core.callbacks.llama_debug import LlamaDebugHandler
 
+from langchain_community.retrievers import TavilySearchAPIRetriever
+from llama_index.core.schema import Document
+from app.utils.search_utils import search_cache, add_documents_with_filtered_metadata
+
+
 # -------------------------------------------------------------------
 # 1) MONKEY-PATCH PARA ChatMessage (por requerimiento de LlamaIndex)
 # -------------------------------------------------------------------
@@ -163,30 +168,43 @@ class GeminiLLM(LLM):
 # -------------------------------------------------------------------
 # 3) HERRAMIENTAS PERSONALIZADAS
 # -------------------------------------------------------------------
+def retrieve_with_fallback(query):
+    cached_results = search_cache.search(query)
+    
+    if cached_results:
+        print(f"✅ Retrieved {len(cached_results)} documents from cache")
+        return format_docs(cached_results)
+    
+    print("🔍 No cache hit, performing web search")
+
+    retriever = TavilySearchAPIRetriever(api_key=os.getenv("TAVILY_API_KEY"), k=5)
+    search_results = retriever.invoke(query)
+    
+    docs = []
+    for result in search_results:
+        if isinstance(result, Document):
+            docs.append(result)
+        else:
+            docs.append(Document(page_content=result))
+    
+    add_documents_with_filtered_metadata(search_cache, docs)
+    
+    return format_docs(docs)
+
+
+
 HEADERS = {'User-Agent': 'Mozilla/5.0'}
 
-def buscar_web(query: str, max_attempts: int = 2, num_results: int = 5) -> str:
-    """
-    Usa DuckDuckGo (vía duckduckgo_search) para devolver hasta 'num_results' resultados.
-    """
-    for i in range(max_attempts):
-        try:
-            with DDGS(headers=HEADERS, timeout=25) as ddgs:
-                results = list(ddgs.text(query, region='es-es', safesearch='moderate', max_results=num_results))
-                if results:
-                    salida = []
-                    for idx, r in enumerate(results):
-                        titulo = r.get('title', 'Sin título')
-                        enlace = r.get('href', 'N/A')
-                        cuerpo = r.get('body', '')
-                        salida.append(f"Fuente {idx+1}: Título: {titulo}\nEnlace: {enlace}\nCuerpo: {cuerpo}")
-                    return "\n\n".join(salida)
-            return "No se encontraron resultados relevantes."
-        except Exception as e:
-            if i < max_attempts - 1:
-                time.sleep(2 * (i + 1))
-            else:
-                return f"Error buscar_web tras {max_attempts} intentos: {e}"
+def buscar_web(query, num_results=5):
+    retriever = TavilySearchAPIRetriever(api_key=os.getenv("TAVILY_API_KEY"), k=num_results)
+    results = retriever.invoke(query)
+    # Formatear resultados para pasarlos al LLM
+    formatted_results = []
+    for i, doc in enumerate(results):
+        formatted_results.append(
+            f"Result {i+1}:\nTitle: {doc.metadata.get('title','')}\nSource: {doc.metadata.get('source','')}\nContent: {doc.page_content}\n"
+        )
+    return "\n\n".join(formatted_results)
 
 def reverse_text(text: str) -> str:
     """Invierte el orden de los caracteres en 'text'."""
@@ -297,10 +315,10 @@ def classify_botanical(items_list_str: str) -> str:
     others_sorted = sorted(set(others))
 
     return (
-        f"Verduras: {', '.join(vegs_sorted)}\n"
-        f"Frutas: {', '.join(fruits_sorted)}\n"
-        f"Otros: {', '.join(others_sorted)}"
-    )
+    f"Vegetables: {', '.join(vegs_sorted)}\n"
+    f"Fruits: {', '.join(fruits_sorted)}\n"
+    f"Others: {', '.join(others_sorted)}"
+)
 
 def scrape_wikipedia_table(page_title: str, section: str, table_index: int = 0) -> str:
     """
@@ -308,7 +326,7 @@ def scrape_wikipedia_table(page_title: str, section: str, table_index: int = 0)
     Devuelve el CSV.
     """
     try:
-        wikipedia.set_lang("es")
+        wikipedia.set_lang("en")
         page = wikipedia.page(page_title, auto_suggest=False)
         soup = BeautifulSoup(page.html(), 'html.parser')
         header = next(
@@ -336,8 +354,9 @@ def scrape_wikipedia_table(page_title: str, section: str, table_index: int = 0)
 search_tool = FunctionTool.from_defaults(
     fn=buscar_web,
     name="web_search",
-    description="Búsqueda DuckDuckGo (máximo 5 resultados)."
+    description="Searches the web using TavilySearch API."
 )
+
 reverse_tool = FunctionTool.from_defaults(
     fn=reverse_text,
     name="reverse_text",
@@ -405,7 +424,7 @@ tool_descriptions = "\n".join([
 # -------------------------------------------------------------------
 # 6) PROMPT DE SISTEMA MEJORADO with few-shot examples
 # -------------------------------------------------------------------
-system_prompt = f"""
+system_prompt_deprecated = f"""
 Eres Alfred, un agente ReAct eficiente y preciso. Tu objetivo es responder correctamente usando las herramientas disponibles.
 A continuación tienes ejemplos de cómo usar cada herramienta:
 
@@ -452,6 +471,65 @@ Herramientas disponibles (USAR EXÁCTAMENTE estos nombres):
 {tool_descriptions}
 """
 
+system_prompt = f"""
+You are Alfred, a ReAct agent. Your goal is to answer correctly using the available tools.
+
+Strict guidelines:
+
+1️. ALWAYS use the available tools first if the question requires information you cannot deduce internally.
+2️. When a tool is used, ONLY answer based on the tool output. DO NOT add or invent any content not explicitly present in the tool output.
+3️. If a tool fails, you may explain the failure clearly. DO NOT fabricate the answer.
+4️. If no tool can help and you don't know, say "I cannot answer with the available tools."
+
+Flow:
+
+- **READ the question carefully.**
+- **SELECT the most appropriate tool:**
+    - `classify_botanical_foods` → grocery list, vegetables, fruits
+    - `read_excel_data` → Excel or attached Excel
+    - `scrape_wiki_table` → Wikipedia, featured articles, tables
+    - `analyze_markdown_table` → Markdown table, commutativity
+    - `reverse_text` → reverse text
+    - `execute_code` → math, code
+    - `web_search` → all other general questions
+- **CALL the tool → COPY its output EXACTLY**
+- **When answering, ONLY use the tool output. DO NOT add any interpretation unless the tool explicitly asked you to process it.**
+
+Few-shot examples:
+
+### Example: classify_botanical_foods
+User: "milk, eggs, broccoli, celery, lettuce"
+Agent: 
+{{
+    "tool": "classify_botanical_foods",
+    "input": "milk, eggs, broccoli, celery, lettuce"
+}}
+Observation: Verduras: broccoli, celery, lettuce
+Frutas: 
+Otros: eggs, milk
+Final Answer: "broccoli, celery, lettuce"
+
+### Example: analyze_markdown_table
+User: "Check commutativity"
+Agent:
+{{
+    "tool": "analyze_markdown_table",
+    "input": "|A|B|C|\\n|---|---|---|\\n|A|A|B|C|..."
+}}
+Observation: a, b
+Final Answer: "a, b"
+
+---
+
+ONLY respond following this flow. DO NOT answer using your internal knowledge if a tool is required and available.
+If unsure, default to using the most appropriate tool first.
+
+Available tools:
+
+{tool_descriptions}
+"""
+
+
 # -------------------------------------------------------------------
 # 7) INICIALIZAR EL AGENTE ReActAgent
 # -------------------------------------------------------------------
@@ -472,7 +550,11 @@ def basic_agent_response(question: str) -> str:
     """
     try:
         if "attached excel" in question.lower() or "archivo excel" in question.lower():
-            return read_excel_data("data/attached.xlsx")
+            excel_result = read_excel_data("data/attached.xlsx")
+            if "Error" in excel_result:
+                return "The Excel file is not available."
+            return excel_result
+            
         resp = alfred_agent.query(question)
         if hasattr(resp, 'response') and resp.response is not None:
             return str(resp.response)
@@ -481,3 +563,4 @@ def basic_agent_response(question: str) -> str:
         return "No se generó una respuesta válida."
     except Exception as e:
         return f"Error crítico del agente: {e}"
+