neograph::CompletionParams Struct Reference

Parameters for an LLM completion request. More...

#include <provider.h>

Public Attributes
std::shared_ptr< graph::CancelToken >	cancel_token
	Optional cancel handle (v0.3+).
json	extra_fields
	Per-call body field bindings (issue #33, v0.8+).
int	max_tokens = -1
	Maximum output tokens. -1 means provider default.
std::vector< ChatMessage >	messages
	Conversation history.
std::string	model
	Model name to use (e.g., "gpt-4o-mini").
float	temperature = 0.7f
	Sampling temperature (0.0 = deterministic, 1.0 = creative).
std::vector< ChatTool >	tools
	Available tools the LLM may call.

Detailed Description

Parameters for an LLM completion request.

Definition at line 57 of file provider.h.

Member Data Documentation

cancel_token

std::shared_ptr<graph::CancelToken> neograph::CompletionParams::cancel_token

Optional cancel handle (v0.3+).

When set, async-native providers (OpenAIProvider, SchemaProvider) bind cancel_token->slot() to their ConnPool::async_post co_await, so a caller's cancel() aborts the in-flight HTTPS socket and stops billable LLM work mid-stream.

The engine populates this from RunConfig::cancel_token when a node calls ctx.provider->complete(params); user code that constructs CompletionParams directly may set it to share an abort across multiple completions.

Definition at line 77 of file provider.h.

extra_fields

json neograph::CompletionParams::extra_fields

Per-call body field bindings (issue #33, v0.8+).

Map of path → value overrides that the provider stamps into the outgoing request body for THIS call only. Companion to the schema-static request.extra_fields block that ships with every body — extra_fields here is the dynamic per-call side of the same hook.

For SchemaProvider: the schema author declares which paths are bindable per-call via "request.per_call_fields": [...]. Only listed paths are honoured; unknown paths are silently dropped (the schema, not the caller, owns the contract). This keeps the per-call surface declarative and matches how temperature_path / max_tokens_path already address specific paths.

Native Provider subclasses (e.g. OpenAIProvider) may choose to honour or ignore individual keys per their own documented surface — the field is generic, the contract is provider-defined.

// Reasoning model — high effort for hard call, low for cheap call.
CompletionParams hard;
hard.extra_fields = {{"reasoning.effort", "high"}};
auto deep = co_await provider->complete_async(hard);
 
CompletionParams cheap;
cheap.extra_fields = {{"reasoning.effort", "low"}};
auto fast = co_await provider->complete_async(cheap);

Default: empty json. Same lifecycle shape as the existing schema-static request.extra_fields block.

Definition at line 115 of file provider.h.

---

The documentation for this struct was generated from the following file:

• include/neograph/provider.h