LlamaIndex

examples/04_llama_index.py demonstrates the LlamaIndex integration — the one that exists because decorating is not always possible. In a real agent the tools are BaseToolSpec methods registered through to_tool_list(), which introspects each method's signature and docstring to build the schema the model sees; putting @agentsight.tool on them means editing every class and risking the agent's own behaviour. The SDK's handler on the framework's dispatcher sees those tool calls without touching them.

python examples/04_llama_index.py
AGENTSIGHT_EXAMPLES_OFFLINE=1 python examples/04_llama_index.py

Offline it uses LlamaIndex's own MockLLM, so the dispatcher, the handler and every span are real — only the model is fake. With OPENAI_API_KEY set it uses a real OpenAI model and installs the OpenAI patch as well, which is the combination a real deployment runs — and the one that proves the rule that keeps the two integrations from double-counting: the handler stands down on LLM spans for any provider a patch already covers, so each call produces exactly one llm span.

What it runs

• A turn with two tool calls and an LLM call — FunctionTool calls the handler observes from the dispatcher, no decorator anywhere. The tools are called directly rather than through an agent loop: an agent has to be persuaded to call a tool, which needs a real model and makes the example about prompt luck instead of instrumentation.
• A failing tool — LlamaIndex may hand the failure back as a ToolOutput rather than raise; either way the span records the error.
• Turnless spend — a completion outside any turn, the background summarisation kind of call that quietly accumulates spend nobody attributes to a user.

What to look for in the output

In examples/traces/04_llama_index/: tool spans that exist despite no @agentsight.tool anywhere in the script, the failed one marked with its error, and — in live mode — one llm span per model call even with two integrations installed.