feat(advisor): Evidence Workspace — structured panes, markdown, sources as knowledge units

Rebuilds the Compliance Advisor floating widget from a plain chat into an Evidence Workspace: pinned last question, markdown-rendered answer (clean prose), and separate panes for Sources (hierarchical Knowledge Units), Figures (C8, conditional) and Footnotes (C-FN), plus a stats bar (Quellen/Regelwerke/Diagramme/Fußnoten). Scrollable turn history; stays a floating icon on every SDK page. Architecture (user direction): the frontend renders ONLY structured evidence and NEVER parses the answer text. The proxy now returns a JSON AdvisorEvidenceMeta line followed by the streamed markdown answer; advisor-rag exposes structured results; an adapter maps RAG/compiler output to the frontend envelope. Figures/footnotes wire in once the RAG-ingestion contract lands (requested on the board) — figures pane is conditional. - lib/sdk/advisor/{evidence,evidence-adapter}.ts (+ adapter test, 7 cases) - components/sdk/advisor/* panes + in-house safe Markdown (no new dep, no dangerouslySetInnerHTML) + test - useAdvisorStream (meta-line parse + streamed answer) + useAdvisorEmail (escaped) - proxy: evidence-meta-v1 envelope + clean-prose prompt (no inline citations) - tsc clean, 11 vitest pass, check-loc 0. ESLint not installed in this node_modules -> CI lints on push. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-07-01 07:46:37 +02:00
parent f0120b237e
commit 49171e841f
22 changed files with 1379 additions and 421 deletions
@@ -0,0 +1,145 @@
+// Adapter: RAG/compiler output -> the structured AdvisorEvidenceMeta the Evidence Workspace renders.
+// This is the ONLY place that maps backend shapes to the frontend envelope. The frontend never
+// parses the answer text — all structure originates here from structured fields.
+
+import type {
+  AdvisorEvidenceMeta,
+  FigureUnit,
+  FootnoteUnit,
+  KnowledgeUnit,
+  RegulationRef,
+} from './evidence'
+import { deriveStats } from './evidence'
+import type { SdkRagResult } from '../agents/advisor-rag'
+
+/** Provisional raw figure (C8) shape — reconcile with the RAG-ingestion contract (board). */
+export interface RawFigure {
+  figure_id?: string
+  id?: string
+  label?: string // "Abbildung 3"
+  caption?: string
+  topic?: string
+  regulation_code?: string
+  regulation_short?: string
+  regulation_name?: string
+  section?: string
+  vision_summary?: string
+  description?: string
+  image_url?: string
+  url?: string
+}
+
+/** Provisional raw footnote (C-FN) shape — reconcile with the RAG-ingestion contract (board). */
+export interface RawFootnote {
+  id?: string
+  ref?: string
+  number?: string | number
+  regulation_code?: string
+  regulation_short?: string
+  regulation_name?: string
+  section?: string
+  text?: string
+}
+
+export interface RawEvidenceInput {
+  results?: SdkRagResult[]
+  figures?: RawFigure[]
+  footnotes?: RawFootnote[]
+}
+
+function regulationRef(
+  code?: string,
+  name?: string,
+  short?: string,
+): RegulationRef {
+  return {
+    code: (code || short || name || 'unknown').toLowerCase().replace(/\s+/g, '_'),
+    name: name || undefined,
+    short: short || name || code || 'Quelle',
+  }
+}
+
+function truncate(text: string, max = 240): string {
+  const t = text.trim().replace(/\s+/g, ' ')
+  return t.length > max ? `${t.slice(0, max - 1)}…` : t
+}
+
+function toKnowledgeUnit(r: SdkRagResult, idx: number): KnowledgeUnit | null {
+  const regulation = regulationRef(r.regulation_code, r.regulation_name, r.regulation_short)
+  const section = r.is_recital
+    ? `Erwägungsgrund ${r.article ?? ''}`.trim()
+    : r.article || undefined
+  const label = r.article_label?.trim() || undefined
+  // Drop empty placeholders: a unit needs at least a label or a section to be meaningful.
+  if (!label && !section && !regulation.name && regulation.short === 'Quelle') return null
+  return {
+    id: `src-${idx}`,
+    regulation,
+    section,
+    paragraph: r.paragraph || undefined,
+    subsection: r.sub || undefined,
+    label,
+    score: typeof r.score === 'number' ? r.score : undefined,
+    snippet: r.text ? truncate(r.text) : undefined,
+    open: r.source_url ? { originalUrl: r.source_url } : undefined,
+  }
+}
+
+function dedupeKey(u: KnowledgeUnit): string {
+  return [u.regulation.code, u.section, u.paragraph, u.subsection, u.label]
+    .map((x) => x || '')
+    .join('|')
+}
+
+function toFigureUnit(f: RawFigure, idx: number): FigureUnit | null {
+  const id = f.figure_id || f.id
+  const imageUrl = f.image_url || f.url
+  if (!id && !imageUrl && !f.label) return null
+  return {
+    id: id || `fig-${idx}`,
+    label: f.label || `Abbildung ${idx + 1}`,
+    caption: f.caption || undefined,
+    topic: f.topic || undefined,
+    source: regulationRef(f.regulation_code, f.regulation_name, f.regulation_short),
+    section: f.section || undefined,
+    visionSummary: f.vision_summary || f.description || undefined,
+    imageUrl: imageUrl || undefined,
+  }
+}
+
+function toFootnoteUnit(f: RawFootnote, idx: number): FootnoteUnit | null {
+  const ref = f.ref || (f.number != null ? `Fußnote ${f.number}` : undefined)
+  if (!ref && !f.text) return null
+  return {
+    id: f.id || `fn-${idx}`,
+    ref: ref || `Fußnote ${idx + 1}`,
+    source: regulationRef(f.regulation_code, f.regulation_name, f.regulation_short),
+    section: f.section || undefined,
+    text: f.text || undefined,
+  }
+}
+
+/**
+ * Build the structured evidence meta. Sources are deduped (same citation retrieved multiple
+ * times collapses to one, keeping the highest score) and order is preserved by score.
+ */
+export function adaptEvidence(input: RawEvidenceInput): AdvisorEvidenceMeta {
+  const seen = new Map<string, KnowledgeUnit>()
+  ;(input.results || []).forEach((r, i) => {
+    const unit = toKnowledgeUnit(r, i)
+    if (!unit) return
+    const key = dedupeKey(unit)
+    const existing = seen.get(key)
+    if (!existing || (unit.score ?? 0) > (existing.score ?? 0)) seen.set(key, unit)
+  })
+  const sources = [...seen.values()].sort((a, b) => (b.score ?? 0) - (a.score ?? 0))
+  const figures = (input.figures || [])
+    .map(toFigureUnit)
+    .filter((x): x is FigureUnit => x !== null)
+  const footnotes = (input.footnotes || [])
+    .map(toFootnoteUnit)
+    .filter((x): x is FootnoteUnit => x !== null)
+
+  const meta = { sources, figures, footnotes }
+  return { ...meta, stats: deriveStats(meta) }
+}
@@ -0,0 +1,103 @@
+// Structured evidence contract for the Compliance Advisor "Evidence Workspace".
+//
+// HARD RULE (architecture): the frontend renders ONLY these structured fields and
+// NEVER parses the answer text. All structure (sources, figures, footnotes) is owned
+// by the SDK/compiler (C-stages) and surfaced as data. The proxy is the adapter that
+// fills this envelope from RAG/compiler output. See memory: advisor-evidence-workspace-no-parse.
+
+/** A regulation / document reference (CRA, EDPB WP248, MaschinenVO, ...). */
+export interface RegulationRef {
+  code: string // canonical id, e.g. "cra", "edpb_wp248", "maschinenvo"
+  name?: string // full name
+  short?: string // short label shown in the card header
+}
+
+/** Openable targets for an evidence item — present only when the SDK can resolve them. */
+export interface OpenTargets {
+  originalUrl?: string // original text / source_url
+  chunkId?: string // retrieved chunk
+  footnoteId?: string // C-FN
+  figureId?: string // C8
+}
+
+/**
+ * A retrieved source as a hierarchical Knowledge Unit, mirroring the compiler:
+ * Regelwerk -> Section (C1/C2) -> Paragraph -> Footnote (C-FN).
+ * Rendered as a card, not a text-list line. E.g. "EDPB WP248 / Kapitel III.B / Fußnote 17".
+ */
+export interface KnowledgeUnit {
+  id: string
+  regulation: RegulationRef
+  section?: string // "Annex I" / "Kapitel III.B" / "Anhang III"
+  subsection?: string // "Abschnitt 2.3"
+  paragraph?: string // Absatz / paragraph
+  footnoteRef?: string // "Fußnote 17" when this unit IS a footnote-backed source
+  label?: string // pre-formatted citation fallback, e.g. "BDSG § 38 Abs. 1"
+  score?: number // retrieval score (optional)
+  snippet?: string // short passage preview (optional) — lets the user peek the cited text
+  open?: OpenTargets
+}
+
+/** A figure (C8) as a Knowledge Unit — never a bare image. Only present when figures exist. */
+export interface FigureUnit {
+  id: string // figure_id
+  label: string // "Abbildung 3"
+  caption?: string // "PDCA-Zyklus"
+  topic?: string
+  source: RegulationRef // "EDPB ..."
+  section?: string
+  visionSummary?: string // vision/LLM description of the figure
+  imageUrl?: string // Playwright PNG; undefined until the RAG-ingestion contract delivers it
+}
+
+/** A footnote (C-FN) as a first-class evidence item. */
+export interface FootnoteUnit {
+  id: string
+  ref: string // "Fußnote 17"
+  source: RegulationRef
+  section?: string
+  text?: string
+}
+
+/** Counts for the stats bar above the answer ("Diese Antwort basiert auf N Quellen"). */
+export interface AdvisorStats {
+  sources: number
+  regulations: number // distinct Regelwerke
+  figures: number
+  footnotes: number
+}
+
+/**
+ * Meta sent by the proxy FIRST (one JSON line), then the answer streams as tokens.
+ * RAG runs before the LLM, so all evidence is known up front and the panes render
+ * immediately while the answer streams in.
+ */
+export interface AdvisorEvidenceMeta {
+  stats: AdvisorStats
+  sources: KnowledgeUnit[]
+  figures: FigureUnit[]
+  footnotes: FootnoteUnit[]
+  relatedDocs?: KnowledgeUnit[]
+}
+
+/** The full evidence a single answer turn holds (meta + the streamed answer markdown). */
+export interface AdvisorEvidence extends AdvisorEvidenceMeta {
+  answer: string // markdown prose, NO inline citations (sources live in the pane)
+}
+
+export function emptyStats(): AdvisorStats {
+  return { sources: 0, regulations: 0, figures: 0, footnotes: 0 }
+}
+
+/** Pure derivation of the stats bar from the evidence items (no parsing of answer text). */
+export function deriveStats(
+  e: Pick<AdvisorEvidenceMeta, 'sources' | 'figures' | 'footnotes'>,
+): AdvisorStats {
+  const regulations = new Set(e.sources.map((s) => s.regulation.code))
+  return {
+    sources: e.sources.length,
+    regulations: regulations.size,
+    figures: e.figures.length,
+    footnotes: e.footnotes.length,
+  }
+}