neograph::llm::SchemaProvider::Config Struct Reference

Configuration for schema-based provider. More...

#include <schema_provider.h>

Public Attributes
std::string	api_key
	API key (overrides env var if set).
std::string	base_url_override
	If non-empty, overrides the schema's connection.base_url. Useful for test doubles and self-hosted OpenAI-compatible endpoints.
std::string	default_model = "gpt-4o-mini"
	Default model name.
bool	prefer_libcurl = false
	Switch the non-streaming HTTP transport to libcurl (HTTP/2.
std::string	schema_path
	Path to schema file, or built-in name ("openai", "claude", "gemini").
int	timeout_seconds = 60
	HTTP request timeout in seconds.
bool	use_websocket = false
	Drive complete_stream over wss:// instead of HTTP/SSE.

Detailed Description

Configuration for schema-based provider.

Definition at line 68 of file schema_provider.h.

Member Data Documentation

prefer_libcurl

bool neograph::llm::SchemaProvider::Config::prefer_libcurl = false

Switch the non-streaming HTTP transport to libcurl (HTTP/2.

• multiplexing + Cloudflare-friendly fingerprint). Default off: empirical bench (dr_compare 2026-04-26) showed parity or slight regression on the LangGraph-equivalent deep-research workload vs the default HTTP/1.1 ConnPool; the multiplex win only materialises when fan-out width dominates per-call latency, and our default config doesn't hit that yet. Flip on if you have parallel-fan-out code hitting Cloudflare-WAF endpoints (where the default httplib path may get fingerprinted out).

Definition at line 95 of file schema_provider.h.

use_websocket

bool neograph::llm::SchemaProvider::Config::use_websocket = false

Drive complete_stream over wss:// instead of HTTP/SSE.

Currently supported only for the "openai-responses" schema — matches OpenAI's WebSocket mode at /v1/responses, which claims ~40% lower latency on agentic rollouts with 20+ tool calls (per developers.openai.com/api/docs/guides/websocket-mode). Throws on complete_stream for any other schema. Has no effect on complete_async / complete() (non-streaming path stays HTTP).

Definition at line 83 of file schema_provider.h.

---

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

• include/neograph/llm/schema_provider.h