Add feedback-aware correction contracts
This commit is contained in:
@@ -9,14 +9,15 @@ import (
|
||||
)
|
||||
|
||||
type StructuredCompletionRequest struct {
|
||||
StageName string `json:"stage_name"`
|
||||
PromptID string `json:"prompt_id,omitempty"`
|
||||
PromptVersion string `json:"prompt_version,omitempty"`
|
||||
ProfileID string `json:"profile_id,omitempty"`
|
||||
SessionID string `json:"session_id,omitempty"`
|
||||
Inputs LLMInputSet `json:"inputs,omitempty"`
|
||||
Vars map[string]any `json:"vars,omitempty"`
|
||||
StructuredOutputRepairAttempts *int `json:"structured_output_repair_attempts,omitempty"`
|
||||
StageName string `json:"stage_name"`
|
||||
PromptID string `json:"prompt_id,omitempty"`
|
||||
PromptVersion string `json:"prompt_version,omitempty"`
|
||||
ProfileID string `json:"profile_id,omitempty"`
|
||||
SessionID string `json:"session_id,omitempty"`
|
||||
Inputs LLMInputSet `json:"inputs,omitempty"`
|
||||
Vars map[string]any `json:"vars,omitempty"`
|
||||
StructuredOutputRepairAttempts *int `json:"structured_output_repair_attempts,omitempty"`
|
||||
Correction *SemanticCorrection `json:"-"`
|
||||
}
|
||||
|
||||
type StructuredCompletionResponse struct {
|
||||
@@ -156,12 +157,14 @@ type ChunkRequest struct {
|
||||
References ReferenceSet `json:"references,omitempty"`
|
||||
LLMProfile string `json:"llm_profile,omitempty"`
|
||||
StructuredOutputRepairAttempts *int `json:"structured_output_repair_attempts,omitempty"`
|
||||
Correction *SemanticCorrection `json:"-"`
|
||||
Metadata map[string]any `json:"metadata,omitempty"`
|
||||
}
|
||||
|
||||
type ChunkPlanResult struct {
|
||||
Plan source.ChunkPlan `json:"plan"`
|
||||
Warnings []Warning `json:"warnings,omitempty"`
|
||||
Plan source.ChunkPlan `json:"plan"`
|
||||
Warnings []Warning `json:"warnings,omitempty"`
|
||||
ModelCandidate *ModelCandidate `json:"-"`
|
||||
}
|
||||
|
||||
type Chunker interface {
|
||||
@@ -283,6 +286,7 @@ type ValidationResult struct {
|
||||
Approved bool `json:"approved"`
|
||||
ReasonCode string `json:"reason_code,omitempty"`
|
||||
Message string `json:"message,omitempty"`
|
||||
CorrectionGuidance string `json:"-"`
|
||||
DiagnosticArtifactPath string `json:"diagnostic_artifact_path,omitempty"`
|
||||
Warnings []Warning `json:"warnings,omitempty"`
|
||||
}
|
||||
|
||||
169
internal/framework/contracts/correction.go
Normal file
169
internal/framework/contracts/correction.go
Normal file
@@ -0,0 +1,169 @@
|
||||
package contracts
|
||||
|
||||
import (
|
||||
"errors"
|
||||
"fmt"
|
||||
"strings"
|
||||
"unicode/utf8"
|
||||
|
||||
"gitea.maximumdirect.net/eric/notarius/internal/core/source"
|
||||
)
|
||||
|
||||
// CorrectionProtocol identifies how a producer can represent the model output
|
||||
// that directly controlled a candidate.
|
||||
type CorrectionProtocol string
|
||||
|
||||
const (
|
||||
CorrectionProtocolSingleResponseV1 CorrectionProtocol = "single_response_v1"
|
||||
)
|
||||
|
||||
const (
|
||||
MaxValidationReasonCodeBytes = 128
|
||||
MaxValidationCorrectionGuidanceBytes = 4 * 1024
|
||||
MaxAssistantResponseBytes = 1 << 20
|
||||
MaxCorrectionGuidanceBytes = 64 * 1024
|
||||
MaxCorrectionContentBytes = MaxAssistantResponseBytes + MaxCorrectionGuidanceBytes
|
||||
)
|
||||
|
||||
func (protocol CorrectionProtocol) Validate() error {
|
||||
if protocol == CorrectionProtocolSingleResponseV1 {
|
||||
return nil
|
||||
}
|
||||
return fmt.Errorf("unsupported correction protocol %q", protocol)
|
||||
}
|
||||
|
||||
// SemanticCorrection carries the latest model response and application-owned
|
||||
// guidance for one fresh corrected request. Its content is sensitive and is
|
||||
// deliberately excluded from ordinary JSON serialization.
|
||||
type SemanticCorrection struct {
|
||||
AssistantResponse []byte `json:"-"`
|
||||
UserGuidance string `json:"-"`
|
||||
}
|
||||
|
||||
func NewSemanticCorrection(assistantResponse []byte, userGuidance string) (*SemanticCorrection, error) {
|
||||
correction := &SemanticCorrection{
|
||||
AssistantResponse: append([]byte(nil), assistantResponse...),
|
||||
UserGuidance: userGuidance,
|
||||
}
|
||||
if err := correction.Validate(); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
return correction, nil
|
||||
}
|
||||
|
||||
func CloneSemanticCorrection(correction *SemanticCorrection) (*SemanticCorrection, error) {
|
||||
if correction == nil {
|
||||
return nil, nil
|
||||
}
|
||||
return NewSemanticCorrection(correction.AssistantResponse, correction.UserGuidance)
|
||||
}
|
||||
|
||||
// CloneStructuredCompletionRequest returns a request whose mutable values are
|
||||
// owned by the caller. It is suitable for clients that retain requests after
|
||||
// CompleteStructured returns.
|
||||
func CloneStructuredCompletionRequest(request StructuredCompletionRequest) (StructuredCompletionRequest, error) {
|
||||
correction, err := CloneSemanticCorrection(request.Correction)
|
||||
if err != nil {
|
||||
return StructuredCompletionRequest{}, fmt.Errorf("clone correction: %w", err)
|
||||
}
|
||||
vars, err := source.CloneMetadata(request.Vars)
|
||||
if err != nil {
|
||||
return StructuredCompletionRequest{}, fmt.Errorf("clone variables: %w", err)
|
||||
}
|
||||
request.Inputs = request.Inputs.Clone()
|
||||
request.Vars = vars
|
||||
request.Correction = correction
|
||||
if request.StructuredOutputRepairAttempts != nil {
|
||||
attempts := *request.StructuredOutputRepairAttempts
|
||||
request.StructuredOutputRepairAttempts = &attempts
|
||||
}
|
||||
return request, nil
|
||||
}
|
||||
|
||||
func (correction SemanticCorrection) Validate() error {
|
||||
if err := validateAssistantResponse(correction.AssistantResponse); err != nil {
|
||||
return fmt.Errorf("semantic correction assistant response: %w", err)
|
||||
}
|
||||
if err := validateBoundedText(correction.UserGuidance, MaxCorrectionGuidanceBytes, "semantic correction user guidance", false); err != nil {
|
||||
return err
|
||||
}
|
||||
if len(correction.AssistantResponse)+len(correction.UserGuidance) > MaxCorrectionContentBytes {
|
||||
return errors.New("semantic correction content exceeds maximum length")
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
// ModelCandidate preserves the exact single model response that directly
|
||||
// controlled a producer result. It is attempt-local and never durable data.
|
||||
type ModelCandidate struct {
|
||||
Response []byte `json:"-"`
|
||||
Protocol CorrectionProtocol `json:"-"`
|
||||
}
|
||||
|
||||
func NewModelCandidate(response []byte, protocol CorrectionProtocol) (*ModelCandidate, error) {
|
||||
candidate := &ModelCandidate{Response: append([]byte(nil), response...), Protocol: protocol}
|
||||
if err := candidate.Validate(); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
return candidate, nil
|
||||
}
|
||||
|
||||
func CloneModelCandidate(candidate *ModelCandidate) (*ModelCandidate, error) {
|
||||
if candidate == nil {
|
||||
return nil, nil
|
||||
}
|
||||
return NewModelCandidate(candidate.Response, candidate.Protocol)
|
||||
}
|
||||
|
||||
func (candidate ModelCandidate) Validate() error {
|
||||
if err := candidate.Protocol.Validate(); err != nil {
|
||||
return err
|
||||
}
|
||||
if err := validateAssistantResponse(candidate.Response); err != nil {
|
||||
return fmt.Errorf("model candidate response: %w", err)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func ValidateValidationResult(result ValidationResult) error {
|
||||
if result.ReasonCode != "" {
|
||||
if err := validateBoundedText(result.ReasonCode, MaxValidationReasonCodeBytes, "validation reason code", false); err != nil {
|
||||
return err
|
||||
}
|
||||
}
|
||||
if result.CorrectionGuidance != "" {
|
||||
if err := validateBoundedText(result.CorrectionGuidance, MaxValidationCorrectionGuidanceBytes, "validation correction guidance", false); err != nil {
|
||||
return err
|
||||
}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func validateAssistantResponse(response []byte) error {
|
||||
if len(response) > MaxAssistantResponseBytes {
|
||||
return errors.New("exceeds maximum length")
|
||||
}
|
||||
if !utf8.Valid(response) {
|
||||
return errors.New("must be valid UTF-8")
|
||||
}
|
||||
if len(strings.TrimSpace(string(response))) == 0 {
|
||||
return errors.New("must not be blank")
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func validateBoundedText(value string, maximum int, name string, optional bool) error {
|
||||
if value == "" && optional {
|
||||
return nil
|
||||
}
|
||||
if !utf8.ValidString(value) {
|
||||
return fmt.Errorf("%s must be valid UTF-8", name)
|
||||
}
|
||||
if strings.TrimSpace(value) == "" {
|
||||
return fmt.Errorf("%s must not be blank", name)
|
||||
}
|
||||
if len(value) > maximum {
|
||||
return fmt.Errorf("%s exceeds maximum length", name)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
141
internal/framework/contracts/correction_test.go
Normal file
141
internal/framework/contracts/correction_test.go
Normal file
@@ -0,0 +1,141 @@
|
||||
package contracts
|
||||
|
||||
import (
|
||||
"bytes"
|
||||
"encoding/json"
|
||||
"strings"
|
||||
"testing"
|
||||
)
|
||||
|
||||
func TestSemanticCorrectionOwnsValidatedContent(t *testing.T) {
|
||||
assistant := []byte(`{"items":["original"]}`)
|
||||
correction, err := NewSemanticCorrection(assistant, "Return one corrected replacement.")
|
||||
if err != nil {
|
||||
t.Fatalf("NewSemanticCorrection() error = %v", err)
|
||||
}
|
||||
assistant[0] = '['
|
||||
if got := string(correction.AssistantResponse); got != `{"items":["original"]}` {
|
||||
t.Fatalf("assistant response = %q, want owned original content", got)
|
||||
}
|
||||
|
||||
clone, err := CloneSemanticCorrection(correction)
|
||||
if err != nil {
|
||||
t.Fatalf("CloneSemanticCorrection() error = %v", err)
|
||||
}
|
||||
clone.AssistantResponse[0] = '['
|
||||
if got := string(correction.AssistantResponse); got != `{"items":["original"]}` {
|
||||
t.Fatalf("source correction changed through clone = %q", got)
|
||||
}
|
||||
if nilClone, err := CloneSemanticCorrection(nil); err != nil || nilClone != nil {
|
||||
t.Fatalf("CloneSemanticCorrection(nil) = %#v, %v; want nil, nil", nilClone, err)
|
||||
}
|
||||
|
||||
encoded, err := json.Marshal(correction)
|
||||
if err != nil {
|
||||
t.Fatalf("marshal correction: %v", err)
|
||||
}
|
||||
if string(encoded) != "{}" {
|
||||
t.Fatalf("correction JSON = %s, want no sensitive content", encoded)
|
||||
}
|
||||
}
|
||||
|
||||
func TestCorrectionContractsRejectInvalidContent(t *testing.T) {
|
||||
tooLongAssistant := bytes.Repeat([]byte("a"), MaxAssistantResponseBytes+1)
|
||||
tooLongGuidance := strings.Repeat("a", MaxCorrectionGuidanceBytes+1)
|
||||
tooLongReason := strings.Repeat("a", MaxValidationReasonCodeBytes+1)
|
||||
tooLongValidationGuidance := strings.Repeat("a", MaxValidationCorrectionGuidanceBytes+1)
|
||||
|
||||
for _, test := range []struct {
|
||||
name string
|
||||
call func() error
|
||||
}{
|
||||
{"blank assistant", func() error { _, err := NewSemanticCorrection([]byte(" \n"), "guidance"); return err }},
|
||||
{"invalid assistant utf8", func() error { _, err := NewSemanticCorrection([]byte{0xff}, "guidance"); return err }},
|
||||
{"oversized assistant", func() error { _, err := NewSemanticCorrection(tooLongAssistant, "guidance"); return err }},
|
||||
{"blank guidance", func() error { _, err := NewSemanticCorrection([]byte("response"), " \t"); return err }},
|
||||
{"invalid guidance utf8", func() error { _, err := NewSemanticCorrection([]byte("response"), string([]byte{0xff})); return err }},
|
||||
{"oversized guidance", func() error { _, err := NewSemanticCorrection([]byte("response"), tooLongGuidance); return err }},
|
||||
{"unsupported protocol", func() error { _, err := NewModelCandidate([]byte("response"), "multiple_responses"); return err }},
|
||||
{"missing candidate protocol", func() error { _, err := NewModelCandidate([]byte("response"), ""); return err }},
|
||||
{"blank candidate response", func() error { _, err := NewModelCandidate([]byte(" "), CorrectionProtocolSingleResponseV1); return err }},
|
||||
{"oversized reason code", func() error { return ValidateValidationResult(ValidationResult{ReasonCode: tooLongReason}) }},
|
||||
{"blank reason code", func() error { return ValidateValidationResult(ValidationResult{ReasonCode: " \t"}) }},
|
||||
{"invalid correction guidance utf8", func() error {
|
||||
return ValidateValidationResult(ValidationResult{CorrectionGuidance: string([]byte{0xff})})
|
||||
}},
|
||||
{"oversized correction guidance", func() error {
|
||||
return ValidateValidationResult(ValidationResult{CorrectionGuidance: tooLongValidationGuidance})
|
||||
}},
|
||||
} {
|
||||
t.Run(test.name, func(t *testing.T) {
|
||||
if err := test.call(); err == nil {
|
||||
t.Fatal("validation error = nil, want error")
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
func TestModelCandidateOwnsValidatedResponse(t *testing.T) {
|
||||
response := []byte(`{"items":["original"]}`)
|
||||
candidate, err := NewModelCandidate(response, CorrectionProtocolSingleResponseV1)
|
||||
if err != nil {
|
||||
t.Fatalf("NewModelCandidate() error = %v", err)
|
||||
}
|
||||
response[0] = '['
|
||||
if got := string(candidate.Response); got != `{"items":["original"]}` {
|
||||
t.Fatalf("candidate response = %q, want owned original content", got)
|
||||
}
|
||||
clone, err := CloneModelCandidate(candidate)
|
||||
if err != nil {
|
||||
t.Fatalf("CloneModelCandidate() error = %v", err)
|
||||
}
|
||||
clone.Response[0] = '['
|
||||
if got := string(candidate.Response); got != `{"items":["original"]}` {
|
||||
t.Fatalf("source candidate changed through clone = %q", got)
|
||||
}
|
||||
if nilClone, err := CloneModelCandidate(nil); err != nil || nilClone != nil {
|
||||
t.Fatalf("CloneModelCandidate(nil) = %#v, %v; want nil, nil", nilClone, err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestValidationResultAllowsAbsentOptionalCorrectionFields(t *testing.T) {
|
||||
if err := ValidateValidationResult(ValidationResult{Approved: true}); err != nil {
|
||||
t.Fatalf("ValidateValidationResult() error = %v, want nil", err)
|
||||
}
|
||||
if err := ValidateValidationResult(ValidationResult{ReasonCode: "invalid-evidence", CorrectionGuidance: "Provide source-backed evidence."}); err != nil {
|
||||
t.Fatalf("ValidateValidationResult() error = %v, want nil", err)
|
||||
}
|
||||
if err := CorrectionProtocol("").Validate(); err == nil {
|
||||
t.Fatal("empty correction protocol validation error = nil, want error")
|
||||
}
|
||||
}
|
||||
|
||||
func TestCloneStructuredCompletionRequestOwnsCorrection(t *testing.T) {
|
||||
correction, err := NewSemanticCorrection([]byte(`{"value":"original"}`), "Correct the value.")
|
||||
if err != nil {
|
||||
t.Fatalf("NewSemanticCorrection() error = %v", err)
|
||||
}
|
||||
attempts := 2
|
||||
request := StructuredCompletionRequest{
|
||||
Inputs: LLMInputSet{"source": NewLLMInputMaterial("source", "application/json", []byte(`{"source":true}`), "", "")},
|
||||
Vars: map[string]any{"labels": []string{"original"}},
|
||||
StructuredOutputRepairAttempts: &attempts,
|
||||
Correction: correction,
|
||||
}
|
||||
clone, err := CloneStructuredCompletionRequest(request)
|
||||
if err != nil {
|
||||
t.Fatalf("CloneStructuredCompletionRequest() error = %v", err)
|
||||
}
|
||||
correction.AssistantResponse[0] = '['
|
||||
request.Inputs["source"] = NewLLMInputMaterial("source", "application/json", []byte(`{"source":false}`), "", "")
|
||||
*request.StructuredOutputRepairAttempts = 7
|
||||
if got := string(clone.Correction.AssistantResponse); got != `{"value":"original"}` {
|
||||
t.Fatalf("cloned correction response = %q, want owned original content", got)
|
||||
}
|
||||
if got := string(clone.Inputs["source"].Content); got != `{"source":true}` {
|
||||
t.Fatalf("cloned input = %q, want owned original content", got)
|
||||
}
|
||||
if clone.StructuredOutputRepairAttempts == nil || *clone.StructuredOutputRepairAttempts != 2 {
|
||||
t.Fatalf("cloned repair attempts = %v, want 2", clone.StructuredOutputRepairAttempts)
|
||||
}
|
||||
}
|
||||
@@ -42,12 +42,14 @@ type TypedExtractionRequest struct {
|
||||
References ReferenceSet
|
||||
LLMProfile string
|
||||
StructuredOutputRepairAttempts *int
|
||||
Correction *SemanticCorrection
|
||||
Metadata map[string]any
|
||||
}
|
||||
|
||||
type TypedExtractionResult[T any] struct {
|
||||
Value T
|
||||
Warnings []Warning
|
||||
Value T
|
||||
Warnings []Warning
|
||||
ModelCandidate *ModelCandidate
|
||||
}
|
||||
|
||||
type Extractor[T any] interface {
|
||||
@@ -65,12 +67,14 @@ type TypedMergeRequest[T any] struct {
|
||||
References ReferenceSet
|
||||
LLMProfile string
|
||||
StructuredOutputRepairAttempts *int
|
||||
Correction *SemanticCorrection
|
||||
Metadata map[string]any
|
||||
}
|
||||
|
||||
type TypedMergeResult[T any] struct {
|
||||
Value T
|
||||
Warnings []Warning
|
||||
Value T
|
||||
Warnings []Warning
|
||||
ModelCandidate *ModelCandidate
|
||||
}
|
||||
|
||||
type Merger[T any] interface {
|
||||
@@ -87,13 +91,15 @@ type TypedNormalizeRequest[T any] struct {
|
||||
References ReferenceSet
|
||||
LLMProfile string
|
||||
StructuredOutputRepairAttempts *int
|
||||
Correction *SemanticCorrection
|
||||
Metadata map[string]any
|
||||
}
|
||||
|
||||
type TypedNormalizeResult[T any] struct {
|
||||
Value T
|
||||
Warnings []Warning
|
||||
Retry *NormalizeRetry
|
||||
Value T
|
||||
Warnings []Warning
|
||||
Retry *NormalizeRetry
|
||||
ModelCandidate *ModelCandidate
|
||||
}
|
||||
|
||||
// Normalize retry diagnostic limits bound module-provided values before the
|
||||
|
||||
Reference in New Issue
Block a user