feat(admin-ui): Operate pages — Query Playground, Ingest, Chat (7.9b)

[officialCodeWork]

Summary

The console can now drive the request path, not just observe it. This closes the three High-priority "no UI yet" gaps from the v2 API-coverage analysis, built to the Operate design brief and 1:1 on the real wire types (rag_core.gateway_types / rag_core.openai_types) — no invented shapes.

Query Playground (/playground, new Operate nav group)

• Full QueryRequest control rail: corpus multiselect, top_k slider, Rerank / Pack / Generate switches, an Advanced disclosure (pack_budget, generate_max_tokens, generate_temperature, key→value filters).
• Run query → POST /v1/query; Retrieve only → POST /v1/retrieve (ChunkRef view).
• Results: routing decision strip (shape · bm25 fallback · corpus strategy · packed tokens · A/B experiment tag), the answer with clickable [n] citation chips + hallucination-guard verdict, citations list, ranked chunks table (trust badges, row → detail sheet), and a stacked per-stage timings bar (skipped stages marked).
• Feedback bar (👍/👎 · 1–5★ · copy · regenerate · PII-redacted comment) → POST /v1/feedback with the exact 9-signal enum; toasts echo the FeedbackAck score.
• View full trace deep-links to /trace?rid=<request_id> (the trace page now accepts rid); the seeded run reuses DEMO_REQUEST_ID, so the demo resolves end-to-end.

Ingest (/ingest, Knowledge group)

• Drag-and-drop multipart POST /v1/ingest/document (10 parser types advertised), staged queue with per-file progress → IngestResult chunk/embedding count badges + "View in Corpora" on success, error + Retry on failure.
• Pipeline explainer: parse → chunk → enrich → PII-redact → embed → index — PII is redacted before embedding.

Chat (/chat)

• OpenAI-compatible playground in three tabs: Chat (streaming SSE thread; right rail = model / temperature / max_tokens / stream + the non-standard rag options, RAG on by default; every reply surfaces the terminal chunk's rag annotations — shape, chunks_n, citations, guard, finish_reason, usage), Embeddings (one input per line, float/base64, dim + L2 norm + sparkline per vector), Models (GET /v1/models — the wired backends, not a catalog).

Plumbing

• Live→seed hybrid as everywhere; these are action surfaces, so without a gateway they simulate (deterministic worked query, token-streamed replies, prefers-reduced-motion honoured).
• New Slider + Tabs primitives; CorpusMultiSelect / GuardBadge / CitedText shared components.
• api.ts: runQuery, retrieveOnly, submitFeedback, ingestDocument (multipart), streamChat (SSE reader → terminal chunk), chatComplete, createEmbeddings, fetchModels. Identity body-carried on /v1/query·/v1/retrieve·/v1/feedback, header-driven elsewhere — matching the gateway.

Verification

• tsc --noEmit clean · next lint clean · 46/46 vitest (12 new across the three pages) · next build — all 17 routes static (the /trace Suspense boundary keeps it prerenderable).
• Browser-verified in light and dark: all three pages, the seeded query → answer → citations → chunks flow, ingest success/failure rows, chat streaming + rag annotations, embeddings cards, and the /trace?rid=… deep link (resolves to the signed demo trace).
• Note: repo CI is currently blocked by the account billing issue (no Actions runs) — gates were run locally.

Documentation

• docs/reference/admin-ui.md — Operate pages, updated page table (16 routes), hybrid/simulation semantics, body-carried identity note, config row.
• TRACKER.md — 7.9b entry under Recently shipped.

🤖 Generated with Claude Code