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

Reuse prepared glossary data across stage sections

Browse Source
This commit is contained in:
Eric Rakestraw
2026-11-10 20:42:28 +00:00
parent 8ad27e8368
commit c16d61e091
21 changed files with 174 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
internal/framework/validators/deterministic.go
View File

@@ -137,7 +137,12 @@ func (v NoEffectValidator) Validate(_ context.Context, req Request) (Result, err
}
func ValidateProtectedTerms(req Request) (Result, error) {
vocab := NewProtectedVocabulary(req.ProtectedTerms)
var vocab ProtectedVocabulary
if req.GlossaryView != nil {
vocab = req.GlossaryView.Protection
} else {
vocab = NewProtectedVocabulary(req.ProtectedTerms)
}
decisions := make([]Decision, 0, len(req.CandidateProposal))
glossaryStage := req.ModuleKey == modulecatalog.KeyGlossary
for _, c := range req.CandidateProposal {

55
internal/framework/validators/llm_context.go
View File

@@ -1,15 +1,43 @@
package validators
import (
"slices"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/chunking"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/glossary"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/schema"
)
// GlossaryView holds the read-only matching data shared by a stage's validators
// and apply-time checks.
type GlossaryView struct {
Protection ProtectedVocabulary
context glossaryContext
}
func PrepareGlossaryView(entries []glossary.Entry, protected []glossary.Term) *GlossaryView {
return &GlossaryView{Protection: NewProtectedVocabulary(protected), context: prepareGlossaryContext(entries)}
}
type glossaryContext struct {
entries []glossary.Entry
positions map[glossary.Identity][]int
vocabulary ProtectedVocabulary
}
func prepareGlossaryContext(entries []glossary.Entry) glossaryContext {
positions := make(map[glossary.Identity][]int, len(entries))
for i, entry := range entries {
identity := glossary.Identity{Collection: entry.Collection, ID: entry.ID}
positions[identity] = append(positions[identity], i)
}
return glossaryContext{entries: entries, positions: positions, vocabulary: NewProtectedVocabulary(glossary.Vocabulary(entries, nil))}
}
// validationContext supplies adjacent source segments within the current chunk
// and glossary definitions mentioned in the candidate or that local context.
// These fields travel in each item so batching accounts for their size.
func validationContext(segments []schema.Segment, position int, selected []glossary.Entry, vocabulary ProtectedVocabulary, corrected string) ([]NeighboringSegment, []glossary.Entry) {
func (view glossaryContext) validationContext(segments []schema.Segment, position int, corrected string) ([]NeighboringSegment, []glossary.Entry) {
var neighbors []NeighboringSegment
source := segments[position].Text
if position > 0 {
@@ -23,19 +51,20 @@ func validationContext(segments []schema.Segment, position int, selected []gloss
for _, neighbor := range neighbors {
texts = append(texts, neighbor.Text)
}
mentioned := make(map[glossary.Identity]struct{})
for _, text := range texts {
for _, occurrence := range view.vocabulary.occurrences(text) {
mentioned[occurrence.identity] = struct{}{}
}
}
var matched []int
for identity := range mentioned {
matched = append(matched, view.positions[identity]...)
}
slices.Sort(matched)
var entries []glossary.Entry
if selected != nil {
mentioned := make(map[glossary.Identity]bool)
for _, text := range texts {
for identity := range vocabulary.occurrencesByIdentity(text) {
mentioned[identity] = true
}
}
for _, entry := range selected {
if mentioned[glossary.Identity{Collection: entry.Collection, ID: entry.ID}] {
entries = append(entries, entry)
}
}
for _, index := range matched {
entries = append(entries, view.entries[index])
}
return neighbors, entries
}

10
internal/framework/validators/llm_validators.go
View File

@@ -9,7 +9,6 @@ import (
"gitea.maximumdirect.net/eric/audita/v3/internal/core/chunking"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/config"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/glossary"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/schema"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/outcomes"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/proposals"
@@ -254,7 +253,12 @@ func BuildLLMValidationRequest(validatorName string, validatorType LLMValidatorT
immediate := make([]Decision, 0)
var segments []schema.Segment
positions := make(map[int]int)
vocabulary := NewProtectedVocabulary(glossary.Vocabulary(req.Glossary, nil))
var glossary glossaryContext
if req.GlossaryView != nil {
glossary = req.GlossaryView.context
} else {
glossary = prepareGlossaryContext(req.Glossary)
}
if req.WorkingTranscript != nil {
segments = req.WorkingTranscript.Segments
for i, seg := range segments {
@@ -276,7 +280,7 @@ func BuildLLMValidationRequest(validatorName string, validatorType LLMValidatorT
continue
}
neighbors, glossaryEntries := validationContext(segments, position, req.Glossary, vocabulary, preview.CorrectedSegmentText)
neighbors, glossaryEntries := glossary.validationContext(segments, position, preview.CorrectedSegmentText)
items = append(items, LLMValidationItem{
CorrectionIndex: p.ProposalIndex,
SegmentID: p.TargetSegmentID,

1
internal/framework/validators/models.go
View File

@@ -45,6 +45,7 @@ type Request struct {
ReplacementPolicy proposals.ReplacementPolicy `json:"replacement_policy"`
Glossary []glossary.Entry `json:"-"`
ProtectedTerms []glossary.Term `json:"-"`
GlossaryView *GlossaryView `json:"-"`
Settings config.Stage `json:"-"`
LLMClient StructuredLLMClient `json:"-"`
DiagnosticsWriter InteractionDiagnosticsWriter `json:"-"`

9
internal/framework/validators/vocabulary_test.go
View File

@@ -110,13 +110,14 @@ func TestExemptGlossarySpellingKeepsLongestMatchBoundary(t *testing.T) {
func TestValidationContextIncludesEachSharedDefinition(t *testing.T) {
entries := []glossary.Entry{
{Collection: "items", ID: "sword", Name: "Sword", Category: "item"},
{Collection: "spells", ID: "shield", Name: "Shield", Category: "spell"},
{Collection: "items", ID: "shield", Name: "Shield", Category: "item"},
{Collection: "items", ID: "sword", Name: "Sword", Category: "item"},
{Collection: "items", ID: "bow", Name: "Bow", Category: "item"},
}
segments := []schema.Segment{{ID: 1, Text: "Shield"}}
_, selected := validationContext(segments, 0, entries, NewProtectedVocabulary(glossary.Vocabulary(entries, nil)), "Barrier")
if len(selected) != 2 || selected[0].Category != "spell" || selected[1].ID != "shield" || selected[1].Collection != "items" {
segments := []schema.Segment{{ID: 1, Text: "Shield Shield"}}
_, selected := prepareGlossaryContext(entries).validationContext(segments, 0, "Sword")
if len(selected) != 3 || selected[0].ID != "sword" || selected[1].Category != "spell" || selected[2].ID != "shield" || selected[2].Collection != "items" {
t.Fatalf("selected=%+v", selected)
}
}