openinference.h File Reference

C++ peer of neograph_engine.openinference (issue #9). More...

#include <neograph/api.h>
#include <neograph/graph/types.h>
#include <neograph/observability/tracer.h>
#include <neograph/provider.h>
#include <memory>
#include <string>

Include dependency graph for openinference.h:

Go to the source code of this file.

Classes
class	neograph::observability::OpenInferenceProvider
	Provider wrapper that emits OpenInference LLM spans. More...
class	neograph::observability::OpenInferenceTracerSession
	RAII session returned by openinference_tracer. More...

Functions
NEOGRAPH_API OpenInferenceTracerSession	neograph::observability::openinference_tracer (Tracer &tracer, std::string root_name="graph.run", std::string node_span_prefix="node.")
	Open a CHAIN-kind root span and return a session whose cb opens per-node CHAIN child spans + LLM_TOKEN events under it.

Detailed Description

C++ peer of neograph_engine.openinference (issue #9).

Two pieces, mirroring the Python module:

• openinference_tracer(tracer) — opens a CHAIN-kind root span, returns an OpenInferenceTracerSession whose cb field plugs into engine.run_stream() / engine.run_stream_async() and opens a CHAIN-kind child span per node, with NODE_START / END payloads stuffed into input.value / output.value JSON blobs and LLM_TOKEN streamed-chunk events recorded as discrete span events. Closing the session ends the root span.
• OpenInferenceProvider(inner, tracer) — wraps any Provider. Each complete* call opens an LLM-kind child span tagged with llm.model_name, llm.invocation_parameters, llm.input_messages.{i}.message.{role,content}, llm.output_messages.0.message.{role,content}, and llm.token_count.{prompt,completion,total}. The streaming overloads (complete_stream / complete_stream_async) also accumulate streamed tokens into the LLM span's output.value and emit one llm.token event per chunk.

Phoenix / Arize / Langfuse render the resulting trace as a chat chain with token counts and per-message bubbles — the same UX the Python wrapper produces.

Tracer adapter: NeoGraph itself does not link against opentelemetry-cpp. Downstream provides an adapter implementing neograph::observability::Tracer that wraps its own backend (OTel SDK, in-memory test fake, logging recorder, etc.). See tests/test_openinference_cpp.cpp for an InMemoryTracer reference impl used by the parity tests.

Definition in file openinference.h.

Function Documentation

openinference_tracer()

NEOGRAPH_API OpenInferenceTracerSession neograph::observability::openinference_tracer	(	Tracer &	tracer,
		std::string	root_name = "graph.run",
		std::string	node_span_prefix = "node."
	)

Open a CHAIN-kind root span and return a session whose cb opens per-node CHAIN child spans + LLM_TOKEN events under it.

Parameters

tracer	Tracer adapter. Must outlive the session.
root_name	Span name for the root (default "graph.run").
node_span_prefix	Prefix for per-node spans (default "node.").