Documentation
¶
Overview ¶
Package topiccontrol implements classify.TopicClassifier against NVIDIA's Llama-3.1-NemoGuard-8B TopicControl model, served by NIM over an OpenAI-compatible chat-completions endpoint.
The model takes a topical instruction as the system message and a conversation whose final entry is the user message to judge, and answers with exactly one of two labels: "on-topic" or "off-topic". Rendering a classify.TopicPolicy into that instruction lives here rather than in the eval handler, so a future non-chat backend can consume the same policy without the pack changing.
Index ¶
Constants ¶
const DefaultModel = "nvidia/llama-3.1-nemoguard-8b-topic-control"
DefaultModel is the TopicControl model id. NIM accepts it as the `model` field on the OpenAI-compatible endpoint it serves.
Variables ¶
This section is empty.
Functions ¶
func ParseLabel ¶
func ParseLabel(raw string) classify.TopicDecision
ParseLabel maps the model's answer onto a decision. Anything that is not one of the two labels is unknown — never silently allow, because "the classifier said something unexpected" and "the message is in scope" are different facts and the policy decides what to do about each.
func RenderPolicy ¶
func RenderPolicy(p classify.TopicPolicy) string
RenderPolicy turns a structured policy into the system instruction the TopicControl model expects.
Types ¶
type Client ¶
type Client struct {
// contains filtered or unexported fields
}
Client implements classify.TopicClassifier against a TopicControl NIM.
func (*Client) ClassifyTopic ¶
func (c *Client) ClassifyTopic( ctx context.Context, req classify.TopicRequest, ) (classify.TopicResult, error)
ClassifyTopic renders the policy as the system instruction, replays history for reference resolution, and puts the message under judgment last.
func (*Client) HTTPTimeout ¶
HTTPTimeout reports the per-call timeout this client will apply. Exported for tests: the timeout is only observable otherwise by waiting for it to expire.
type Config ¶
type Config struct {
// BaseURL is the OpenAI-compatible root of the NIM deployment, e.g.
// http://topic-control:8000/v1. Required — there is no public default
// endpoint for a self-hosted NIM, and guessing one would turn a
// misconfiguration into a confusing timeout.
BaseURL string
// APIKey is optional: a locally deployed NIM usually needs none, while
// NVIDIA-hosted endpoints do.
APIKey string
// Model overrides DefaultModel.
Model string
// Timeout bounds a single classification call. Zero means
// defaultHTTPTimeout. Note this caps the call regardless of the deadline
// on the context passed to ClassifyTopic: whichever is shorter wins, so a
// caller cannot extend it by supplying a longer context.
Timeout time.Duration
// HTTPClient overrides the default client entirely, Timeout included.
// Mainly for tests.
HTTPClient *http.Client
}
Config configures a Client.
Source Files