Configure a Custom AI Model · Documentation

Before PacketSafari configuration

Serve the model through an OpenAI-compatible endpoint with Responses API streaming and automatic tool calling enabled. Record the exact inference-server and model revisions.

Configure the parser required by the installed inference-server release. Current vLLM examples include:

Model family	Example vLLM tool parser
Kimi K2	`kimi_k2`
Kimi K3	`kimi_k3` tool and reasoning parsers

The installed vLLM or SGLang documentation is authoritative. Parser names and behavior can change between releases, and the parser configuration is part of the qualification fingerprint.

DeepSeek V4 Flash support applies to the tested 0731 behavior family. Do not reuse that result for an earlier Flash revision or a different inference-server translation layer; qualify the exact customer-operated route.

Kimi K3 is an always-reasoning model. Preserve the complete assistant reasoning and tool-call history across continuation turns. The public 8× B300 option is a reference profile; final configuration and sizing are confirmed during deployment design.

PacketSafari workflow

Open Admin → AI settings.
Select the matching provider, or select an OpenAI-compatible provider for a customer inference gateway.
Enter the base URL and authentication mode, then save.
If the endpoint is outside the PacketSafari container network, complete egress approval.
Use Test connection, then Refresh models. If discovery is unavailable, enter the exact served model ID manually.
Under Runtime defaults → Manual model onboarding, add or import a model profile. Record the context window, maximum output, reasoning behavior, tool parser, structured-output support, and server/model revisions.
Map the model to the intended default or strongest role and save.
Select the model and run Qualify selected model. Confirm the three short synthetic calls. No packet capture is sent.
Run a small representative PacketSafari investigation before production use. Confirm the answer, exact packet evidence, runtime, and tool behavior.

Do not apply OpenAI-style reasoning_effort or verbosity settings unless the exact endpoint declares and passes those parameters. Always preserve reasoning and tool history when the model or server requires it.

Organization-specific endpoints use Settings → Organization → Organization AI. The same connection, discovery, profile, qualification, and egress rules apply, and organization settings take precedence for their members.