Represents built-in tools available from AI/LLM services that can be used within the LangChain framework.

Native tools are functionalities provided directly by language model services (like Google AI, OpenAI) that can be invoked by LLMs to perform specific actions or retrieve information. They are "native" because they're built into the AI service itself, rather than implemented externally.

Each native tool has:

• A unique name that identifies it (e.g., "google_search", "code_interpreter")
• A configuration map that contains tool-specific settings

Google Search Grounding

Starting with Gemini 2.0, Google Search is available as a native tool. This enables the model to decide when to use Google Search to enhance the factuality and recency of responses.

Google Search as a tool enables:

• More accurate and up-to-date answers with grounding sources
• Retrieving information from the web for further analysis
• Finding relevant images, videos, or other media for multimodal reasoning
• Supporting specialized tasks like coding, technical troubleshooting, etc.

Example with Google Search

alias LangChain.Chains.LLMChain
alias LangChain.Message
alias LangChain.NativeTool

model = ChatGoogleAI.new!(%{temperature: 0, stream: false, model: "gemini-2.0-flash"})

{:ok, updated_chain} =
   %{llm: model, verbose: false, stream: false}
   |> LLMChain.new!()
   |> LLMChain.add_message(
     Message.new_user!("What is the current Google stock price?")
   )
   |> LLMChain.add_tools(NativeTool.new!(%{name: "google_search", configuration: %{}}))
   |> LLMChain.run()

Tool Search

tool_search/1 builds a provider-neutral search tool. Add it to a chain alongside functions marked defer_loading: true (see "Tool search options" in LangChain.Function). Deferred functions are declared on every request but kept out of the model's context until the model searches for them, so the request prefix never changes and prompt caching keeps working.

chain
|> LLMChain.add_tools(NativeTool.tool_search())
|> LLMChain.add_tools(deferred_functions)

Keep at least one function, normally the most-used ones, not deferred.

A chat model that offers tool search translates the tool for its provider. A chat model without tool search falls back: it leaves the search tool out of the request and sends deferred functions as ordinary tools, so every tool is available from the start. Search items another model added to the conversation are left out, and :tool_reference parts are sent as text. A chain configured for tool search therefore runs on any chat model, which keeps fallbacks and model swaps working.

"Live" means a model searched for a deferred tool, called it, and accepted the replayed conversation on a later turn. "Request encoding" means unit tests of the request body only.

A chat model without tool search uses drop_tool_search/1 to leave the search tool out.