LlamaIndex

Available · pip install "agent-adapter[llamaindex]" · llama-index-core>=0.14,<0.15

Every registry tool becomes a LlamaIndex FunctionTool with fn_schema=spec.args_model. Both call paths run through the executor: acall directly, call through run_sync, so a sync agent works inside a running event loop too. LlamaIndex is imported only when you export a tool.

Functions

Function (agent_adapter.integrations.llamaindex)	Returns
llamaindex_tool(spec, *, executor=None)	one FunctionTool
llamaindex_tools(registry, *, executor=None)	one per tool whose targets include llamaindex
INTEGRATION	the integration object

executor=None uses the default executor.

Run

from agent_adapter.integrations.llamaindex import llamaindex_tools

tools = llamaindex_tools(registry, executor=Executor())  # pass these to a LlamaIndex agent
tool = tools[0]
output = await tool.acall(item="tea")  # what an async agent does
sync_output = tool.call(item="tea", region="eu")  # sync agents: run_sync bridge
bad = await tool.acall(item=3)  # invalid args come back as an error value

flowchart LR
  REG[("Registry")] --> FT["FunctionTool<br/>fn_schema = args_model"]
  FT -- "acall(**kwargs)" --> EX["Executor.run_spec"]
  FT -- "call(**kwargs)" --> RS["run_sync"] --> EX
  EX --> OUT["ToolOutput<br/>raw_output = value or error"]

What the model sees

metadata.get_parameters_dict() is the args model's JSON schema (only type, properties, required and $defs at the root); the description is the tool's own. to_openai_tool() works as usual. Tools imported from a raw schema (spec_from_schema) show their loose model's schema, so nested objects appear as plain object / array.

Approval

LlamaIndex has no approval gate or argument-edit primitive of its own. The executor's approval port decides (Executor(approval=...)), and the executor applies a decision's edited_args (validated again). A denial comes back as an approval-denied error value.

Use a port that answers in the call: CallbackApproval (ask a reviewer in-process), AutoApprove or AutoDeny. DeferredApproval cannot resume here: LlamaIndex gives the tool no call id of its own, so a re-run is a new call with a new id, and the ApprovalPending it raises does not pause the agent (see below).

Ids

LlamaIndex passes no call id to a tool, so ctx.call_id is a fresh call_… id and each call is its own executor run.

Errors to expect

Case	ToolOutput / what you get
ToolArgsInvalid, ToolError, ApprovalDenied, GuardrailBlocked, RunLimitExceeded	raw_output = {"error": {"kind", "message"}}
any other error (ToolTimeout, ToolFailed, ...)	raised from acall / call
asyncio.CancelledError	propagates from acall

Inside a LlamaIndex agent workflow, an exception from a tool (including ApprovalPending, ToolTimeout, ToolFailed) is caught by the agent and sent back to the model as a ToolOutput(is_error=True) with the error text; the run goes on.