docs: consistency-review fixes (env-var names, defaults, stale claims, fabricated content)
Consistency review: 5 parallel reviewers verified every concrete doc claim against code. This MR fixes the CRITICAL/MAJOR findings — wrong env-var names, wrong defaults, stale claims, fabricated content, broken links.
Highest-impact fixes (operators would have been misled)
-
rag retrieval/pipeline: every retrieval knob var name was wrong (
RETRIEVE_K→ realRETRIEVER_ARANGO_K); copying them into.envwould have silently no-op'd.TRAVERSAL_MAX_DEPTH2→1; addedRETRIEVER_HYBRID_RRF_K. -
rag reranking:
RERANKER_STRATEGY→RERANKING_STRATEGY; added missingslice_threshold;THRESHOLD0.9→0.75;TOP_N2→1; calibration default isnone(sigmoid is opt-in — naively enabling compresses scores). -
rag generation: LLM default
granite-4.1-8b→ real compose defaultMeta-Llama-3.1-8B-Instruct(granite-4.1-8b = recommended). -
gpu.md:
VLLM_MAX_MODEL_LENT4 65536→2048, RTX6000→4096 (matched env.t4/env.rtx6000); GPU_UTIL reconciliation. -
install-guide: deleted a public
VLLM_MODEL_ID"to be removed" TODO row; stale model ids → real;Admin/ADMINadmin→genie-admin/GENIE_ADMIN_PASSWORD; retriever port 7025→7000. - integration-architecture: "observability not implemented (planned ELK)" → the real OTel+Victoria stack; wrong models (Granite-3.3-2b / BGE-M3) → real defaults.
- mobile-architecture: rewrote a fabricated "Proxy" service layer (ChatbotProxy/ChatHistoryProxy/api_service.dart — none exist) to the real Riverpod-providers + services structure; split KeycloakConfig vs flavor configs.
-
observability/overview:
@tracing.trace_spandecorator (doesn't exist) → FastAPI auto-instrumentation +with_span; 10→9 dashboards; receiver names corrected. -
operations/scaling:
VLLM_GPU_UTIL→VLLM_GPU_UTILIZATION.
Also: dashboard-count fixes, "Story X.Y" BMAD refs removed, source-tree filename typo, mobile _index missing entry + locale/token counts, api-contracts broken ../.claude links removed, dev-guide app.js → createApp().
Build verified clean (0 errors).
Follow-ups (not in this MR — additive/larger rewrites)
- api-contracts-backend.md: enumerate
/admin/queries/inspect+/:queryIdroutes (doc incomplete, not wrong). - source-tree-analysis.md: Vue CLI vs Vite contradiction; top-level tree completeness.
- mobile-architecture.md §6: version numbers vs pubspec.yaml.
- integration-architecture.md: remove "Document History" scan artefact;
ARANGO_USERhardcoded in snippet. - cross-page: layer-stack drawn 3× (project-overview/integration/architecture) — dedup; keycloak admin-user + cert matrix duplicated across compose/swarm/install.
- audience-framing intros missing on state-management / theme-system / mobile-architecture.
- install-guide.md: Confluence-conversion cleanup (backslash escapes, label lines), clone-URL consistency, proofreading.