Rework the Compliance Advisor header ("Diese Antwort stuetzt sich auf")
to describe the EVIDENCE rather than the documents: binding
Rechtsgrundlagen split from Leitlinien (soft-law guidance), a
per-regulation breakdown, plus Abbildungen, Fussnoten and Evidence Units.
No fabricated trust score — objective counts only.
- bindingness is a canonical Legal-KG fact (APEX rule): added an optional
EvidenceUnit.bindingness contract seam; the FE renders the split from it
and degrades to a neutral per-regulation breakdown when it is absent
(SDK/RAG asked via board to populate it in /retrieve).
- evidence-grouping.ts: pure, tested grouping/counting model.
- route.ts: optional `audience` field (tonality) kept out of the retrieval
question; answers lead with a "Kurz gesagt" summary, structured by theme.
- E2E + unit tests updated for the evidence framing.
Not deployed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The advisor endpoint now serves two shapes off one orchestration:
- new FE ({question}) -> v3 JSON contract (clarity/answer/evidence/citations/...).
- legacy consumer ({message}, e.g. breakpilot-workspace which reads a text stream and
persists raw bytes) -> plain-text stream of the L2 answer (clean prose, no [n] markup,
no clarify gate). isLegacyRequest() discriminates; answerSystem() gains withCitations.
Prevents the v3 contract from breaking breakpilot-workspace's chat (CLAUDE.md rule #4,
keep every consumer working). No deploy. tsc clean, 13 vitest (incl. isLegacyRequest),
check-loc 0.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Completes the advisor stack (FE + orchestration; /retrieve is SDK/RAG-owned). The route
now returns the FE contract instead of a text stream:
- retrieveFull() calls /retrieve with {query, context}; consumes clarity/evidence/
visual_evidence/footnotes (exact shape per board 2026-07-01 12:25).
- mode-routing (resolveMode): clarify unless a context was chosen and /retrieve's
clarity.mode says so. clarify -> L1 general answer (completeAdvisorAnswer, ungrounded,
no sources). answer -> L2 answer over numbered evidence with [n] markers.
- citations generated here ([n] -> nth evidence unit); footnotes remapped; evidence /
visual_evidence passed through.
- advisor-llm: non-streaming completeAdvisorAnswer(). Pure mappings in retrieve-mapping.ts
(+ tests). Removed the dead v2 evidence.ts/evidence-adapter (RegulationRef moved to
regulation-display). controls-augmentation kept (tested; re-integrable later).
NOT deployed: joint deploy with the SDK /retrieve endpoint (deploy-coupling). tsc clean,
25 vitest (mapping/clarify/answer/markdown/registry/rag), check-loc 0.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Two structural fixes (not query-specific):
- Proxy prompt: forbid ANY trailing "Quellen:"/"Quellen im RAG-System" list and make it
the LAST instruction so it overrides the soul file's answer-structure + example that
teach a closing sources section. Applies to every answer.
- KnowledgeUnitCard: render the label only when it differs from regulation.short, so a
source whose label == short name no longer prints twice. Applies to every source.
Answer text is still never parsed in the FE (sources live in the pane). + card test.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
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>
Authority-/Freshness-Layer Punkte 1/2/5 im Advisor-Antwortpfad (Prompt-Ebene, kein
Schema). Neue Soul-Sektion "Korpus-Autoritaet & Aktualitaet": rechtliche FAKTEN
(Schwellen/Fristen/Zahlen/Pflichten) nur aus bereitgestelltem RAG-/Controls-Kontext,
Trainingswissen nie als Rechtsquelle; Konflikt -> Kontext gewinnt, transparent;
Co-Pilot-Ton statt Roboter-Verweigerung. Ergaenzt Quellentreue (Fundstellen) um die
Fakten-Ebene -> loest den "DSB ab 10 statt 20"-Fall. route.ts: RAG-Framing als
"deine EINZIGEN Rechtsquellen" verschaerft.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Der Floating-Compliance-Advisor war auf prod kaputt (502): RAG ging ueber
rag-service:8097 (auf prod nicht vorhanden) und der Chat ueber
OLLAMA_URL=ollama-embed (embedding-only, kein qwen2.5vl).
- RAG laeuft jetzt ueber die ai-compliance-sdk /sdk/v1/rag/search (bge-m3,
prod-erreichbar) statt rag-service -> profitiert vom reicheren Embedding.
(lib/sdk/agents/advisor-rag.ts)
- LLM-Kaskade: OVH/LiteLLM (gpt-oss-120b) zuerst, Ollama als Dev-Fallback.
(lib/sdk/agents/advisor-llm.ts; OVH-Env via orca-infra admin-Block)
- ai-sdk: bp_compliance_recht in AllowedCollections ergaenzt (Whitelist war
inkonsistent — die Fehlermeldung listete es bereits als erlaubt).
- Route auf die Module umgestellt (duenn); Controls-Augmentation unveraendert.
- Tests: advisor-rag + advisor-llm.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Prompt-augments the RAG-only advisor with the shared use-case->controls API:
deterministic topic detection -> local controls API -> context block, so the
agent can answer from real Control-IDs. 100% local at runtime (no Anthropic).
NOT pushed/deployed: the shared API currently returns MASTER-grain controls,
whose composition is broken (gpre2 object-only clustering -> mega-clusters).
Pending the atom-grain rework of the API. tsc + vitest green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ollama entlädt das 35b-Modell nach 5 Min Leerlauf → jede Frage danach
startet es kalt (Modell-Load) und läuft in den Frontend-Timeout ("Load
failed"). keep_alive='30m' im Chat-Request hält es warm.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Compliance Advisor, Drafting Agent und Validator haben nicht geantwortet
weil qwen3.5 standardmaessig im Thinking-Mode laeuft (interne Chain-of-
Thought > 2min Timeout). Keiner der Agenten benoetigt Thinking-Mode —
alle Aufgaben sind Chat/Textgenerierung/JSON-Validierung ohne tiefes
Reasoning. think:false sorgt fuer direkte schnelle Antworten.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Increase num_predict from 2048 to 8192 to prevent mid-sentence cutoff
- Add "Quellenschutz" rules to system prompt: agent refuses to list all
available sources/collections, only reveals sources used in answers
- Remove internal collection names from RAG context sent to LLM
- Agent confirms knowledge on specific topics but refuses meta-queries
like "what sources do you have?"
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Ingestion script: Add 3 new PDFs (IFRS DE/EN, EFRAG Endorsement Status)
to ingest-industry-compliance.sh (7 → 10 documents total)
- System prompt: Add EU-IFRS and EFRAG to competence area, add mandatory
IFRS endorsement warning section for all IFRS/IAS queries
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Replace single DSFA corpus query with parallel search across 6 collections
via RAG service (port 8097)
- Add country parameter with metadata filter for bp_compliance_gesetze
- Add country-specific system prompt section
- Add DE/AT/CH/EU toggle buttons in ComplianceAdvisorWidget header
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Reduce chat history from 10 to 6 messages to fit context window
- Lower num_predict from 8192 to 2048 for faster responses
- Add Training module link to SDK sidebar navigation
- Add snake_case to camelCase key transformation for reporting API
(Go backend returns snake_case, TypeScript expects camelCase)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Benjamin Admin