Update helper.py

helper.py

@@ -47,51 +47,86 @@ repl_tool = Tool(
 )
 
 
-def get_travily_api_search_tool(tavily_api_key: str) -> Tool:
+from langchain_community.tools import TavilySearchResults
+from langchain.tools import Tool
+from typing import List, Optional
+
+def get_travily_api_search_tool(
+    tavily_api_key: str,
+    # Optional parameters for more fine-grained control
+    search_depth: str = "advanced",
+    include_domains: Optional[List[str]] = None,
+    exclude_domains: Optional[List[str]] = None,
+    time_range: Optional[str] = None
+) -> Tool:
     """
-    Creates and returns a LangChain Tool for Tavily Search API.
+    Creates and returns a LangChain Tool for Tavily Search API, optimized for GAIA-like
+    fact-based questions to improve consistency and accuracy.
 
     Args:
         tavily_api_key: The API key for Tavily Search.
+        search_depth: 'basic' or 'advanced'. 'advanced' can provide more comprehensive results.
+                      Default: 'advanced' for better factual depth.
+        include_domains: Optional list of domains to specifically include in search results.
+                         Useful for targeting highly reputable sources.
+        exclude_domains: Optional list of domains to specifically exclude from search results.
+                         Useful for filtering out unreliable sources.
+        time_range: Optional time range for results (e.g., 'day', 'week', 'month', 'year').
+                    Useful for stabilizing results for historical facts or focusing on recent data.
 
     Returns:
         A LangChain Tool configured for Tavily Search.
     """
     tavily_search = TavilySearchResults(
-        max_results=3,             # Limits the number of search results to 3
-        include_answer=True,       # Prioritizes direct answers from search results
-        include_raw_content=False, # Excludes full raw content of pages to save tokens
-        include_images=False,      # Excludes image results to save tokens
-        tavily_api_key=tavily_api_key
+        max_results=5,             # Increased from 3 to 5 (or more) for a larger pool of data.
+                                   # This increases the chance of finding the correct answer even if rankings shift.
+        include_answer=False,      # **CRITICAL CHANGE for GAIA L1/L2:**
+                                   # For precise factual answers, it's generally better to
+                                   # let YOUR agent's LLM synthesize the answer from raw snippets.
+                                   # Tavily's `include_answer` uses its own LLM which adds another
+                                   # layer of potential variability and might not always be perfectly aligned
+                                   # with GAIA's strict answer format.
+        include_raw_content=False, # Good for saving tokens, keep as is.
+        include_images=False,      # Good for saving tokens, keep as is.
+        tavily_api_key=tavily_api_key,
+        # Incorporating new parameters for better control:
+        search_depth=search_depth, # Passed from function argument, default 'advanced'
+        include_domains=include_domains,
+        exclude_domains=exclude_domains,
+        time_range=time_range
     )
 
+    # Updated description to reflect changes and emphasize strict input/output for GAIA
     return Tool(
         name="tavily_search",
         description="""
-        A powerful search engine tool for real-time information retrieval from the web.
+        A powerful and precise search engine tool for real-time, factual information retrieval from the web.
+        Optimized for fact-based questions (like GAIA Level 1 & 2).
+
         Use this tool when you need to:
-        - Find up-to-date information, facts, or statistics.
-        - Research current events, people, or concepts.
+        - Find definitive, up-to-date facts, statistics, or direct answers.
+        - Research specific historical events, scientific data, or definitions.
         - Answer questions that require external knowledge not present in the model's training data.
-        - Verify information or get factual answers.
+        - Verify information or get unambiguous factual answers.
 
         **Input Format (CRITICAL):**
-        The input MUST be a concise, clear, and specific search query string.
-        Think of what you would type into a Google search bar.
-        Example: "current weather in London"
-        Example: "population of India 2024"
-        Example: "history of the internet"
+        The input MUST be a concise, clear, and highly specific search query string.
+        Formulate your query as if you're trying to get a single, factual answer from a search engine.
+        Example: "population of Tokyo 2023"
+        Example: "date of birth of Marie Curie"
+        Example: "chemical formula of water"
+        Example: "winner of 2022 FIFA World Cup"
 
         **DO NOT:**
-        - Ask natural language questions that are not search queries.
-        - Provide incomplete sentences or ambiguous terms.
-        - Include personal information or sensitive data in your queries.
-        - Expect direct execution of commands or calculations; this is a search tool.
+        - Ask natural language questions that are not optimized for search queries (e.g., "Tell me a story about...").
+        - Provide incomplete sentences, ambiguous terms, or conversational filler.
+        - Include personal information or sensitive data.
+        - Expect direct execution of commands or calculations; this is solely a search and information retrieval tool.
 
         **Output:**
-        The tool returns a JSON string containing relevant search results.
-        It prioritizes a direct 'answer' if available. Otherwise, it provides a list of
-        search results including titles, URLs, and snippets.
+        The tool returns a JSON string containing relevant search results, primarily focused on providing
+        snippets and URLs to help you extract the precise factual answer.
+        It does NOT provide a direct generated answer but raw search results for your interpretation.
         """,
         func=tavily_search.run,
     )