Skip to content

Agent tools

index.tools() returns tools for an AI agent. Each tool is a plain function plus a JSON schema, so it works with any framework or API.

tools = index.tools(principal="user:ana")
by_name = {tool.name: tool for tool in tools}
Tool Arguments Returns
search_documents query, limit, collection the best chunks, each with a citation
get_context query, budget the best passages, cited, within a budget
list_documents collection the documents and their sizes
outline_document reference the sections, with numbers and sizes
read_document reference, budget a document, a section or pages

Tools return text, which models read best. Errors come back as text too (Error: ...), so the model can correct itself.

With an API

response = client.responses.create(
    model=model,
    input=messages,
    tools=[tool.to_openai() for tool in tools],  # .to_openai_chat() for Chat Completions
)
for call in response.output:
    if call.type == "function_call":
        output = by_name[call.name].call_json(call.arguments)

message = client.messages.create(
    model=model,
    max_tokens=2048,
    messages=messages,
    tools=[tool.to_anthropic() for tool in tools],
)
for block in message.content:
    if block.type == "tool_use":
        output = by_name[block.name].call_json(block.input)

for tool in tools:
    print(tool.name, tool.parameters)
tool = by_name["get_context"]
print(tool(query="refund policy", budget=1000))

The principal is bound, never asked

principal is fixed when the tools are made. It is not a tool argument, so the model can't ask for documents its user may not see. Make the tools once per user (or per request). See Access rules.