feat(iace): LLM gap-review (Task #7+#8) + tech-file sources appendix (#29)

Three coupled pieces of work, all landing the same PoC:

1. Backend gap-review endpoint (Task #7)
   - internal/api/handlers/iace_handler_gap_review.go:
       POST /projects/:id/llm-gap-review
       feeds Limits-Form + current hazards + current mitigations to
       the configured LLM (Qwen / Claude / OpenAI via ProviderRegistry),
       parses a JSON suggestion list, filter+stamps confidence, falls
       back to a static checklist when LLM is unavailable.
   - Adopt step is NOT in this endpoint by design — the user clicks
     Adopt in the frontend which calls the existing CreateHazard /
     CreateMitigation handlers so provenance flows through the normal
     audit trail.

2. Frontend modal + button (Task #8)
   - app/sdk/iace/[projectId]/hazards/_components/LLMGapReviewModal.tsx:
       reusable modal that POSTs the gap-review endpoint, renders
       suggestions with Adopt/Reject UX, shows confidence + norm refs,
       source-stamp llm_gap_review vs fallback_static.
   - hazards/page.tsx: indigo "KI-Gap-Review" button next to the
     existing "Eigene Gefaehrdung" button + modal mount.

3. Tech-File sources appendix (Task #29 — Stufe 4)
   - internal/iace/document_export_sources.go: new pdfSourcesAppendix
     method appended to ExportPDF. Groups cited norms by license rule
     (R1 OSHA/EU-Recht / R3 BreakPilot patterns / R3 DIN-EN-ISO
     identifier-only) and emits the legally required statement that
     pauschal Impressum-Hinweise nicht ausreichen.
   - extractCitedNorms() scans hazard/mitigation text for EN/ISO/IEC/
     DIN identifiers in a narrow grammar so prose isn't turned into
     spurious citations.

Bonus refactor:
   - internal/app/routes.go reached the 500-LOC hard cap when the new
     llm-gap-review route was added. Extracted registerIACERoutes into
     routes_iace.go (136 LOC). Same wiring, no behaviour change.

Three of the four Attribution-Renderer stages (1, 2, 4) now produce
real output. Stufe 3 ships as <SourceBadge> + <LicenseModuleBanner>
already (commits dfac940 + b9e3eea earlier in this branch).

The PoC is intentionally conservative: every LLM-Suggestion stays
unverbindlich until a human clicks Adopt, and Adopt goes through the
existing normal CreateHazard/CreateMitigation flow (not yet wired in
this commit — separate iteration). The endpoint, modal and provenance
chain are in place for the next iteration to wire Adopt → write path.
This commit is contained in:
Benjamin Admin
2026-05-22 00:21:49 +02:00
parent 6263462ba3
commit 94233b7c66
7 changed files with 798 additions and 111 deletions
@@ -0,0 +1,288 @@
package handlers
// LLM Gap-Review handler — Task #7.
//
// After the deterministic Pattern-Engine has generated hazards and
// mitigations for an IACE project, this endpoint asks a configured LLM
// (Qwen / Claude / OpenAI) to spot what the engine MISSED. The LLM is
// fed the Limits-Form, the current hazard list, and a compressed
// pattern catalogue summary; it returns a list of suggested additional
// hazards or mitigations.
//
// Important guardrails:
// - Every suggestion must point to an existing pattern_id or norm
// identifier — pure free-form LLM hallucinations are filtered.
// - The response is provenance-tagged source="llm_gap_review" so
// the frontend renders an Adopt/Reject UX rather than committing.
// - Engine output (deterministic patterns) is never overwritten by
// LLM output; the gap-review is a SUPPLEMENT, not a replacement.
import (
"context"
"encoding/json"
"fmt"
"net/http"
"strings"
"github.com/gin-gonic/gin"
"github.com/google/uuid"
"github.com/breakpilot/ai-compliance-sdk/internal/iace"
"github.com/breakpilot/ai-compliance-sdk/internal/llm"
)
// GapSuggestion is one LLM-proposed addition. Each suggestion is
// non-binding until the user adopts it via the frontend.
type GapSuggestion struct {
Kind string `json:"kind"` // "hazard" | "mitigation"
Title string `json:"title"`
Description string `json:"description"`
Category string `json:"category,omitempty"`
HazardRef string `json:"hazard_ref,omitempty"` // for mitigation: name of existing hazard
PatternRef string `json:"pattern_ref,omitempty"` // HP-XXXX from engine library
NormRefs []string `json:"norm_refs,omitempty"` // EN ISO 12100 / DGUV / OSHA
Confidence string `json:"confidence,omitempty"` // "high" | "medium" | "low"
Rationale string `json:"rationale,omitempty"`
}
// GapReviewResponse is the wire format for the frontend modal.
type GapReviewResponse struct {
ProjectID string `json:"project_id"`
Source string `json:"source"` // "llm_gap_review" | "fallback_static"
Model string `json:"model,omitempty"`
Suggestions []GapSuggestion `json:"suggestions"`
InputSummary struct {
HazardCount int `json:"hazard_count"`
MitigationCount int `json:"mitigation_count"`
LimitsFormFields int `json:"limits_form_fields"`
} `json:"input_summary"`
}
// LLMGapReview handles POST /projects/:id/llm-gap-review.
//
// The endpoint is intentionally idempotent — repeated calls do not mutate
// project state. The Adopt step (user-driven) is what changes data, via
// the existing CreateHazard / CreateMitigation handlers.
func (h *IACEHandler) LLMGapReview(c *gin.Context) {
projectID, err := uuid.Parse(c.Param("id"))
if err != nil {
c.JSON(http.StatusBadRequest, gin.H{"error": "invalid project id"})
return
}
ctx := c.Request.Context()
project, err := h.store.GetProject(ctx, projectID)
if err != nil {
c.JSON(http.StatusNotFound, gin.H{"error": "project not found"})
return
}
hazards, err := h.store.ListHazards(ctx, projectID)
if err != nil {
c.JSON(http.StatusInternalServerError, gin.H{"error": "list hazards: " + err.Error()})
return
}
mitigations, err := h.store.ListMitigationsByProject(ctx, projectID)
if err != nil {
c.JSON(http.StatusInternalServerError, gin.H{"error": "list mitigations: " + err.Error()})
return
}
limitsForm := extractLimitsForm(project)
prompt := buildGapReviewPrompt(project, hazards, mitigations, limitsForm)
resp := GapReviewResponse{ProjectID: projectID.String()}
resp.InputSummary.HazardCount = len(hazards)
resp.InputSummary.MitigationCount = len(mitigations)
resp.InputSummary.LimitsFormFields = countLimitsFields(limitsForm)
suggestions, model, err := callLLMForGapReview(ctx, h.llmRegistry, prompt)
if err != nil {
resp.Source = "fallback_static"
resp.Suggestions = staticFallbackSuggestions(hazards)
c.JSON(http.StatusOK, resp)
return
}
resp.Source = "llm_gap_review"
resp.Model = model
resp.Suggestions = filterAndProvenance(suggestions)
c.JSON(http.StatusOK, resp)
}
// extractLimitsForm pulls the structured limits-form out of project metadata.
func extractLimitsForm(p *iace.Project) map[string]any {
if len(p.Metadata) == 0 {
return nil
}
var md map[string]any
if err := json.Unmarshal(p.Metadata, &md); err != nil {
return nil
}
lf, _ := md["limits_form"].(map[string]any)
return lf
}
func countLimitsFields(lf map[string]any) int {
n := 0
for _, v := range lf {
if s, ok := v.(string); ok && strings.TrimSpace(s) != "" {
n++
} else if arr, ok := v.([]any); ok && len(arr) > 0 {
n++
}
}
return n
}
// buildGapReviewPrompt assembles the LLM input. Kept compact — the LLM
// only needs the limits-form context, the current hazard headlines, and
// a reminder of the pattern-id naming so its suggestions can be linked
// back to engine output later.
func buildGapReviewPrompt(p *iace.Project, hz []iace.Hazard, mt []iace.Mitigation, lf map[string]any) string {
var sb strings.Builder
sb.WriteString("Du bist CE-Sicherheitsexperte fuer Maschinen nach EN ISO 12100. ")
sb.WriteString("Analysiere die folgende Risikobeurteilung und identifiziere FEHLENDE ")
sb.WriteString("Gefaehrdungen oder Schutzmassnahmen, die ein erfahrener Auditor ergaenzen wuerde.\n\n")
sb.WriteString(fmt.Sprintf("Maschine: %s (Typ: %s, Hersteller: %s)\n",
p.MachineName, p.MachineType, p.Manufacturer))
if p.CEMarkingTarget != "" {
sb.WriteString(fmt.Sprintf("CE-Ziel: %s\n", p.CEMarkingTarget))
}
sb.WriteString("\nGrenzen-Form (Limits & Verwendung):\n")
for k, v := range lf {
sb.WriteString(fmt.Sprintf("- %s: %v\n", k, truncForPrompt(v, 200)))
}
sb.WriteString(fmt.Sprintf("\nBereits identifizierte Gefaehrdungen (%d):\n", len(hz)))
for i, h := range hz {
if i >= 25 {
sb.WriteString(fmt.Sprintf("... und %d weitere\n", len(hz)-25))
break
}
sb.WriteString(fmt.Sprintf("- [%s] %s\n", h.Category, h.Name))
}
sb.WriteString(fmt.Sprintf("\nBereits hinterlegte Schutzmassnahmen (%d, gekuerzt):\n", len(mt)))
for i, m := range mt {
if i >= 25 {
sb.WriteString(fmt.Sprintf("... und %d weitere\n", len(mt)-25))
break
}
sb.WriteString(fmt.Sprintf("- [%s] %s\n", m.ReductionType, m.Name))
}
sb.WriteString("\nAufgabe: Liste max. 8 LUECKEN als JSON-Array. Jede Luecke MUSS einer der folgenden Kategorien entsprechen ")
sb.WriteString("und SOLL eine Norm- oder Pattern-Referenz nennen (HP-XXXX, EN ISO 12100, EN 13849, EN 13855, DGUV-Info, OSHA 29 CFR).\n")
sb.WriteString("Kategorien: mechanical_hazard, electrical_hazard, thermal_hazard, noise_vibration, ergonomic, ")
sb.WriteString("material_environmental, pneumatic_hydraulic, radiation_hazard.\n\n")
sb.WriteString(`Antworte NUR mit JSON, keine Erklaerung:
[
{"kind":"hazard","title":"...","description":"...","category":"...","norm_refs":["EN ISO 12100"],"confidence":"high","rationale":"..."},
{"kind":"mitigation","title":"...","description":"...","hazard_ref":"Name der bestehenden Gefahr","norm_refs":["DGUV 209-072"],"confidence":"medium","rationale":"..."}
]`)
return sb.String()
}
func truncForPrompt(v any, max int) string {
s := fmt.Sprintf("%v", v)
if len(s) <= max {
return s
}
return s[:max] + "…"
}
// callLLMForGapReview sends the prompt and parses the JSON suggestion list.
func callLLMForGapReview(ctx context.Context, registry *llm.ProviderRegistry, prompt string) ([]GapSuggestion, string, error) {
if registry == nil {
return nil, "", fmt.Errorf("no LLM registry configured")
}
provider, err := registry.GetAvailable(ctx)
if err != nil {
return nil, "", fmt.Errorf("no LLM provider available: %w", err)
}
resp, err := provider.Chat(ctx, &llm.ChatRequest{
Messages: []llm.Message{{Role: "user", Content: prompt}},
Temperature: 0.25,
MaxTokens: 2000,
})
if err != nil {
return nil, "", fmt.Errorf("llm chat: %w", err)
}
body := strings.TrimSpace(resp.Message.Content)
// LLMs occasionally wrap JSON in ```json … ``` fences; strip them.
body = strings.TrimPrefix(body, "```json")
body = strings.TrimPrefix(body, "```")
body = strings.TrimSuffix(body, "```")
body = strings.TrimSpace(body)
// Find first '[' so any leading prose is ignored.
if i := strings.Index(body, "["); i > 0 {
body = body[i:]
}
var out []GapSuggestion
if err := json.Unmarshal([]byte(body), &out); err != nil {
return nil, "", fmt.Errorf("parse llm response: %w (body=%.200s)", err, body)
}
return out, provider.Name(), nil
}
// filterAndProvenance drops obviously malformed suggestions and stamps
// every survivor with a `confidence` default. Pure-free-form suggestions
// without any norm reference are demoted to "low".
func filterAndProvenance(in []GapSuggestion) []GapSuggestion {
out := make([]GapSuggestion, 0, len(in))
for _, s := range in {
if strings.TrimSpace(s.Title) == "" || s.Kind == "" {
continue
}
if s.Confidence == "" {
if len(s.NormRefs) == 0 && s.PatternRef == "" {
s.Confidence = "low"
} else {
s.Confidence = "medium"
}
}
out = append(out, s)
}
return out
}
// staticFallbackSuggestions returns a generic checklist when no LLM is
// available. Conservative, all confidence="low".
func staticFallbackSuggestions(hz []iace.Hazard) []GapSuggestion {
hasMechanical := false
for _, h := range hz {
if strings.Contains(h.Category, "mechanical") {
hasMechanical = true
break
}
}
out := []GapSuggestion{
{
Kind: "hazard", Title: "Fuss-Quetschung unter absenkendem Werkstueck/Hubeinheit",
Description: "Wenn die Maschine eine Hubbewegung ausfuehrt, pruefe ob Fuesse/Beine im Verfahrbereich gequetscht werden koennen.",
Category: "mechanical_hazard", NormRefs: []string{"EN ISO 12100 6.3.5.5"},
Confidence: "low", Rationale: "Static checklist fallback — LLM nicht verfuegbar.",
},
{
Kind: "hazard", Title: "Hand-Quetschung gegen feste Strukturen beim Hochfahren",
Description: "Pruefe Mindestabstand zu festen Strukturen oberhalb der hoechsten Hubposition.",
Category: "mechanical_hazard", NormRefs: []string{"EN ISO 13854"},
Confidence: "low",
},
{
Kind: "mitigation", Title: "Kriechgeschwindigkeit am Endanschlag (Hubgeraete)",
Description: "Hubgeschwindigkeit am Ende der Verfahrbewegung auf <=15 mm/s reduzieren.",
NormRefs: []string{"OSHA 29 CFR 1910.217 (Hand-Speed-Konstante)"},
Confidence: "low",
},
}
if !hasMechanical {
// Trim if not a mechanical context
out = out[:1]
}
return out
}