internal/llm: add interface PolicyChecker

A PolicyChecker is used to check the inputs and outputs
of an LLM against safety policies. We will implement this
interface with the GCP Checks Guardrails API.

For #70

Change-Id: I52a776cc94900cef4c0b56f284c56e44f5136d4b
Reviewed-on: https://go-review.googlesource.com/c/oscar/+/637975
Reviewed-by: Hyang-Ah Hana Kim <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>

Author: tatianab

internal/llm/policy_checker.go

@@ -0,0 +1,124 @@
+// Copyright 2024 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package llm
+
+import (
+	"context"
+	"fmt"
+)
+
+// A PolicyChecker checks inputs and outputs to LLMs against
+// safety policies.
+type PolicyChecker interface {
+	// SetPolicies sets the policies to evaluate in subsequent
+	// calls to [Check]. If unset, use the implementation's default.
+	SetPolicies([]*PolicyConfig)
+	// CheckText evaluates the policies configured on this [PolicyChecker]
+	// against the given text and returns a result for each [PolicyConfig].
+	// If the text represents a model output, the prompt parts used to generate it
+	// may optionally be provided as context. If the text represents a model input,
+	// prompt should be empty.
+	CheckText(ctx context.Context, text string, prompt ...Part) ([]*PolicyResult, error)
+}
+
+// A PolicyConfig is a policy to apply to an input or output to an LLM.
+//
+// Copied from "google.golang.org/api/checks/v1alpha" to avoid direct dependency.
+type PolicyConfig struct {
+	// PolicyType: Required. Type of the policy.
+	PolicyType PolicyType
+	// Threshold: Optional. Score threshold to use when deciding if the content is
+	// violative or non-violative. If not specified, the default 0.5 threshold for
+	// the policy will be used.
+	Threshold float64
+}
+
+// A PolicyResult is the result of evaluating a policy against
+// an input or output to an LLM.
+//
+// Copied from "google.golang.org/api/checks/v1alpha" to avoid direct dependency.
+type PolicyResult struct {
+	// PolicyType: Type of the policy.
+	PolicyType PolicyType
+	// Score: Final score for the results of this policy.
+	Score float64
+	// ViolationResult: Result of the classification for the policy.
+	ViolationResult ViolationResult
+}
+
+type PolicyType string
+
+// Possible values for [PolicyType].
+const (
+	// Default.
+	PolicyTypeUnspecified = PolicyType("POLICY_TYPE_UNSPECIFIED")
+	// The model facilitates, promotes or enables access to
+	// harmful goods, services, and activities.
+	PolicyTypeDangerousContent = PolicyType("DANGEROUS_CONTENT")
+	// The model reveals an individual’s personal
+	// information and data.
+	PolicyTypePIISolicitingReciting = PolicyType("PII_SOLICITING_RECITING")
+	// The model generates content that is malicious,
+	// intimidating, bullying, or abusive towards another individual.
+	PolicyTypeHarassment = PolicyType("HARASSMENT")
+	// The model generates content that is sexually
+	// explicit in nature.
+	PolicyTypeSexuallyExplicit = PolicyType("SEXUALLY_EXPLICIT")
+	// The model promotes violence, hatred, discrimination on the
+	// basis of race, religion, etc.
+	PolicyTypeHateSpeech = PolicyType("HATE_SPEECH")
+	// The model provides or offers to facilitate access to
+	// medical advice or guidance.
+	PolicyTypeMedicalInfo = PolicyType("MEDICAL_INFO")
+	// The model generates content that contains
+	// gratuitous, realistic descriptions of violence or gore.
+	PolicyTypeViolenceAndGore = PolicyType("VIOLENCE_AND_GORE")
+	// The model generates profanity and obscenities.
+	PolicyTypeObscenityAndProfanity = PolicyType("OBSCENITY_AND_PROFANITY")
+)
+
+// AllPolicyTypes returns a policy that, when passed to
+// to [PolicyChecker.SetPolicies], configures the PolicyChecker
+// to check for all available dangerous content types at the default threshold.
+func AllPolicyTypes() []*PolicyConfig {
+	return []*PolicyConfig{
+		{PolicyType: PolicyTypeDangerousContent},
+		{PolicyType: PolicyTypePIISolicitingReciting},
+		{PolicyType: PolicyTypeHarassment},
+		{PolicyType: PolicyTypeSexuallyExplicit},
+		{PolicyType: PolicyTypeHateSpeech},
+		{PolicyType: PolicyTypeMedicalInfo},
+		{PolicyType: PolicyTypeViolenceAndGore},
+		{PolicyType: PolicyTypeObscenityAndProfanity},
+	}
+}
+
+type ViolationResult string
+
+// Possible values for [ViolationResult].
+const (
+	// Unspecified result.
+	ViolationResultUnspecified = ViolationResult("VIOLATION_RESULT_UNSPECIFIED")
+	// The final score is greater or equal the input score
+	// threshold.
+	ViolationResultViolative = ViolationResult("VIOLATIVE")
+	// The final score is smaller than the input score
+	// threshold.
+	ViolationResultNonViolative = ViolationResult("NON_VIOLATIVE")
+	// There was an error and the violation result could
+	// not be determined.
+	ViolationResultClassificationError = ViolationResult("CLASSIFICATION_ERROR")
+)
+
+// IsViolative reports whether the policy result represents
+// a violated policy.
+func (pr *PolicyResult) IsViolative() bool {
+	return pr.ViolationResult == ViolationResultViolative
+}
+
+// String returns a string representation of the policy result.
+func (pr *PolicyResult) String() string {
+	return fmt.Sprintf("%s: %s (%f)", pr.PolicyType, pr.ViolationResult, pr.Score)
+}