LlamaIndex

Native callback support is on the roadmap. The pattern below works today with any LlamaIndex query engine.

Pattern

Wrap query engine calls in a telemetry execution. Use MODEL_INVOCATION for the LLM call and TOOL_CALL for retrievers.

import fact0
from llama_index.core import VectorStoreIndex

client = fact0.Client(api_key="f0_live_...")
index = VectorStoreIndex.from_documents(documents)
query_engine = index.as_query_engine()

with client.telemetry.execution(agent_id="rag-bot") as ex:
    with ex.span("retrieve", span_type="TOOL_CALL") as span:
        nodes = query_engine.retrieve("What is Fact0?")
        span.complete(output={"hits": len(nodes)})

    with ex.span("synthesize", span_type="MODEL_INVOCATION") as span:
        response = query_engine.query("What is Fact0?")
        span.complete(output={"answer": str(response)})