eric/audita
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

Simplify references and output instructions in LLM prompts

Browse Source
This commit is contained in:
Eric Rakestraw
2026-09-25 12:38:23 +00:00
parent 325c4ea7d3
commit 011b021d47
31 changed files with 248 additions and 129 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
internal/framework/proposal_generation/generate_test.go
View File

@@ -12,6 +12,7 @@ import (
ck "gitea.maximumdirect.net/eric/ConfigKit"
"gitea.maximumdirect.net/eric/audita/internal/core/config"
"gitea.maximumdirect.net/eric/audita/internal/core/schema"
"gitea.maximumdirect.net/eric/audita/internal/framework/contracts"
"gitea.maximumdirect.net/eric/audita/internal/framework/outcomes"
"gitea.maximumdirect.net/eric/audita/internal/framework/proposals"
@@ -77,6 +78,7 @@ func defaultRequest(t *testing.T) Request {
PromptID: "modules.grammar.proposal",
PromptVersion: "v1",
Variables: map[string]string{"SectionJSON": "[]"},
WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 7, Text: "teh"}, {ID: 25, Text: "other"}}},
StartIndex: 0,
}
}
@@ -86,7 +88,7 @@ func TestGenerateCandidatesSuccess(t *testing.T) {
responses: []StructuredCorrectionSet{
{
Corrections: []StructuredCorrectionProposal{
{TargetSegmentID: 7, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9},
{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9},
},
},
},
@@ -104,6 +106,9 @@ func TestGenerateCandidatesSuccess(t *testing.T) {
if len(got.Corrections) != 1 || len(got.Enriched) != 1 {
t.Fatalf("unexpected proposal lengths: %+v", got)
}
if got.Corrections[0].TargetSegmentID != 7 {
t.Fatalf("local reference not resolved: %+v", got.Corrections)
}
if got.Enriched[0].ProposalIndex != 10 {
t.Fatalf("expected proposal index 10, got %d", got.Enriched[0].ProposalIndex)
}
@@ -118,7 +123,7 @@ func TestGenerateCandidatesSuccess(t *testing.T) {
func TestGenerateCandidatesUsesCorrectionSetSchema(t *testing.T) {
client := &fakeStructuredClient{
responses: []StructuredCorrectionSet{
{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
},
}
req := defaultRequest(t)
@@ -140,7 +145,7 @@ func TestGenerateCandidatesUsesCorrectionSetSchema(t *testing.T) {
func TestGenerateCandidatesDiagnosticsIncludeSchemaMetadata(t *testing.T) {
client := &fakeStructuredClient{
responses: []StructuredCorrectionSet{
{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "teh", CorrectedText: "the", Confidence: 0.9}}},
},
}
diag := &captureDiagnosticsWriter{}
@@ -172,7 +177,7 @@ func TestGenerateCandidatesInvalidCorrectionIsPreservedForLaterValidation(t *tes
responses: []StructuredCorrectionSet{
{
Corrections: []StructuredCorrectionProposal{
{TargetSegmentID: 0, OriginalText: "x", CorrectedText: "", Confidence: 1.2},
{SegmentReference: 0, OriginalText: "x", CorrectedText: "", Confidence: 1.2},
},
},
},
@@ -235,8 +240,8 @@ func TestGenerateCandidatesValidEmptyResponseIsSuccessful(t *testing.T) {
func TestGenerateCandidatesDeterministicIndexAssignment(t *testing.T) {
baseResponse := StructuredCorrectionSet{
Corrections: []StructuredCorrectionProposal{
{TargetSegmentID: 1, OriginalText: "a", CorrectedText: "A", Confidence: 0.9},
{TargetSegmentID: 2, OriginalText: "b", CorrectedText: "B", Confidence: 0.9},
{SegmentReference: 1, OriginalText: "a", CorrectedText: "A", Confidence: 0.9},
{SegmentReference: 2, OriginalText: "b", CorrectedText: "B", Confidence: 0.9},
},
}
clientA := &fakeStructuredClient{responses: []StructuredCorrectionSet{baseResponse}}
@@ -266,8 +271,8 @@ func TestGenerateCandidatesDeterministicIndexAssignment(t *testing.T) {
func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
client := &fakeStructuredClient{
responses: []StructuredCorrectionSet{
{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: "alpha", CorrectedText: "ALPHA", Confidence: 0.9}}},
{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 3, OriginalText: "charlie", CorrectedText: "CHARLIE", Confidence: 0.9}}},
{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "alpha", CorrectedText: "ALPHA", Confidence: 0.9}}},
{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: "charlie", CorrectedText: "CHARLIE", Confidence: 0.9}}},
},
}
section0 := contracts.SectionMetadata{Index: 0}
@@ -276,6 +281,7 @@ func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
req0 := defaultRequest(t)
req0.LLMClient = client
req0.Section = &section0
req0.WorkingTranscript = &schema.Transcript{Segments: []schema.Segment{{ID: 71, Text: "alpha"}}}
req0.StartIndex = 0
part0, err := GenerateCandidates(context.Background(), req0)
if err != nil {
@@ -285,6 +291,7 @@ func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
req1 := defaultRequest(t)
req1.LLMClient = client
req1.Section = &section1
req1.WorkingTranscript = &schema.Transcript{Segments: []schema.Segment{{ID: 903, Text: "charlie"}}}
req1.StartIndex = len(part0.Enriched)
part1, err := GenerateCandidates(context.Background(), req1)
if err != nil {
@@ -295,6 +302,9 @@ func TestGenerateCandidatesMultipleSectionsStableMetadata(t *testing.T) {
if len(all) != 2 {
t.Fatalf("expected 2 proposals, got %d", len(all))
}
if all[0].TargetSegmentID != 71 || all[1].TargetSegmentID != 903 {
t.Fatalf("section references did not reset: %+v", all)
}
if all[0].ProposalIndex != 0 || all[1].ProposalIndex != 1 {
t.Fatalf("unexpected proposal indexes: %d, %d", all[0].ProposalIndex, all[1].ProposalIndex)
}
@@ -308,7 +318,7 @@ func TestGenerateCandidatesDiagnosticsWrittenAndRedacted(t *testing.T) {
validationSecret := "proposal-validation-secret"
client := &fakeStructuredClient{
responses: []StructuredCorrectionSet{
{Corrections: []StructuredCorrectionProposal{{TargetSegmentID: 1, OriginalText: validationSecret, CorrectedText: "safe", Confidence: 0.9}}},
{Corrections: []StructuredCorrectionProposal{{SegmentReference: 1, OriginalText: validationSecret, CorrectedText: "safe", Confidence: 0.9}}},
},
}
t.Setenv("PRIMARY_SECRET", primarySecret)