refactor: port tabbed Playground/Dashboard UI from zeabur-rag

- Merge Chat, Report, Learn into tabbed index.html
- Add Inspect, Triage, Edit tabs to Playground
- Simplify nav to [Playground | Dashboard] pill
- Move Settings/Refresh to tab bar as icon-only buttons
- Add tooltip help on stat cards (Triage, Dashboard, Query Signals)
- Add source color legend in Query Signals
- Fix formatAuditChanges to read old_value/new_value
- Rename "Similarity" to "Score" with mode-aware tooltips
- Redirect /report and /learn to /?tab=xxx
- Remove standalone learn.html and report.html
- Add zeabur-rag-feedback skill
- Sync skill doc fixes (RRF vs cosine explanations)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: canyugs

skills/zeabur-rag-curate/SKILL.md

@@ -15,7 +15,7 @@ Auth: `Authorization: Bearer $RAG_API_KEY` — **admin scope required**
 ## Loop
 
 1. **Call `zeabur-rag-triage`** to get the pending items list.
-2. **Present results** as a numbered list grouped by category (🔴 Reports, 🟡 Unverified, 🟠 Low-similarity, 🔵 Negative feedback). Show counts.
+2. **Present results** as a numbered list grouped by category (🔴 Reports, 🟡 Unverified, 🟠 Low-score, 🔵 Negative feedback, ⚪ Needs-frontmatter). Show counts.
 3. **If all categories are empty:** tell the user "The knowledge base has no pending items — looks healthy ✅" and stop.
 4. **Wait for user** to pick an item by number or description.
 5. **Run the decision tree** for that item type (see below).
@@ -51,7 +51,7 @@ Auth: `Authorization: Bearer $RAG_API_KEY` — **admin scope required**
 | Correct but needs polish | `zeabur-rag-edit` to fix → then verify |
 | Low quality or wrong | Reject (see Inline commands) |
 
-### 🟠 Low-similarity signal
+### 🟠 Low-score signal
 
 1. Call `zeabur-rag-search` with the original query to confirm the gap.
 2. Classify:
@@ -75,6 +75,30 @@ Auth: `Authorization: Bearer $RAG_API_KEY` — **admin scope required**
 | Answer was right but user query was ambiguous | Dismiss — no action |
 | No chunks to inspect (null top_chunk_ids) | Treat as knowledge gap → `zeabur-rag-learn` |
 
+### ⚪ Needs-frontmatter
+
+A chunk was ingested from a source file that has no frontmatter. Its ID is a
+temporary path-hash (`MAN-tmp-*`), and it is excluded from some reorg-safety
+guarantees. The fix is to add frontmatter to the source file in the origin
+repo (e.g. `zebra-manual`), then re-ingest.
+
+1. Call `zeabur-rag-inspect` to show the chunk's `url` and content.
+2. Tell the user which file to edit (the `url` field points at the repo-relative
+   path, e.g. `docs/customer-support/xxx.md`).
+3. Direct them to the backfill tool in the origin repo. For `zebra-manual`, that
+   tool lives in `zebra-manual/scripts/` and walks through one file at a time
+   with LLM-drafted frontmatter + human confirmation.
+4. After the user finishes adding frontmatter, tell them to re-run the pipeline:
+   `bun run pipeline --adapter zebra-manual --input /path/to/zebra-manual/docs`
+5. The adapter's reverse-lookup will automatically delete the old `tmp-*` chunks
+   and insert the new permanent-ID chunks in place. No manual DB surgery.
+
+| Classification | Action |
+|---------------|--------|
+| File is real and worth keeping | User adds frontmatter in origin repo → re-ingest |
+| File is not worth keeping in KB | Reject the chunk (see Inline commands); user also deletes the source file |
+| File is a duplicate of another chunk | Reject this chunk; point user at the canonical one |
+
 ## Inline commands
 
 These actions don't have dedicated skills. Execute them as direct API calls:
@@ -86,7 +110,7 @@ curl -s -X POST "$ZEABUR_RAG_URL/api/admin/learned/<id>/verify" \
   -H "Authorization: Bearer $RAG_API_KEY"
 ```
 
-Returns `{"success": true}`. The chunk now ranks at full weight in search (unverified `learned` chunks are downranked to ×0.7 similarity).
+Returns `{"success": true}`. The chunk now ranks at full weight in search (unverified `learned` chunks are downranked to ×0.7 score).
 
 ### Reject a learned chunk (soft delete)
 

skills/zeabur-rag-feedback/SKILL.md

@@ -0,0 +1,52 @@
+---
+name: zeabur-rag-feedback
+description: Send feedback on RAG search results — thumbs-up (+1) or thumbs-down (-1) to indicate whether the answer was helpful. Use after calling zeabur-rag-search when you can evaluate the quality of the results. Positive feedback reinforces good results; negative feedback flags poor matches for review. Requires write:feedback scope.
+---
+
+# RAG Feedback
+
+Send feedback on search results to improve knowledge base quality over time.
+
+Base URL: `$ZEABUR_RAG_URL`
+Auth: `Authorization: Bearer $RAG_API_KEY`
+
+## API
+
+```bash
+curl -s -X POST "$ZEABUR_RAG_URL/api/feedback" \
+  -H "Authorization: Bearer $RAG_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "query": "how to deploy to zeabur",
+    "answer": "the generated answer text",
+    "chunks": ["DOC-1234", "SUP-5678"],
+    "score": 1,
+    "signal_id": "optional-signal-id",
+    "comment": "optional explanation"
+  }'
+```
+
+| Field | Required | Description |
+|-------|----------|-------------|
+| `query` | Yes | The original search query |
+| `answer` | No | The generated RAG answer |
+| `chunks` | No | Array of chunk IDs that were returned |
+| `score` | Yes | `1` (helpful) or `-1` (not helpful) |
+| `signal_id` | No | Signal ID from the search response — links feedback to the specific query signal |
+| `comment` | No | Free-text explanation of why the result was good or bad |
+
+## When to send feedback
+
+| Situation | Score |
+|-----------|-------|
+| Answer correctly addressed the question | `1` |
+| Retrieved chunks were relevant and accurate | `1` |
+| Answer was wrong, incomplete, or misleading | `-1` |
+| Retrieved chunks were irrelevant to the query | `-1` |
+| No useful results were returned | `-1` |
+
+## Notes
+
+- If `signal_id` is provided, the feedback score is written back to the `rag_query_signals` table, linking it to the original query for analytics.
+- Negative feedback surfaces in the Triage review queue for admin follow-up.
+- Feedback is also appended to `data/user-feedback.jsonl` as a backup log.

skills/zeabur-rag-inspect/SKILL.md

@@ -42,7 +42,7 @@ curl -s "$ZEABUR_RAG_URL/api/admin/chunks/<chunk_id>" \
     { "id": "...", "type": "outdated", "query": "...", "detail": "...", "status": "open", "created_at": "..." }
   ],
   "signals": [
-    { "id": "...", "query": "...", "mode": "hybrid", "top_similarity": 0.45, "feedback_score": null, "created_at": "..." }
+    { "id": "...", "query": "...", "mode": "hybrid", "top_similarity": 0.012, "feedback_score": null, "created_at": "..." }
   ],
   "audit_log": [
     { "id": "...", "chunk_id": "SUP-1234", "action": "edit", "old_value": {}, "new_value": {}, "created_at": "..." }

skills/zeabur-rag-search/SKILL.md

@@ -34,7 +34,7 @@ curl -s -X POST "$ZEABUR_RAG_URL/api/query" \
 
 **Always pass `client: "claude-code"` when calling from Claude Code** so queries are trackable in the dashboard.
 
-Response includes `chunks[]` (each with `id`, `title`, `answer`, `tags`, `similarity`, `source`, `verified`) and `signal_id` for linking feedback.
+Response includes `chunks[]` (each with `id`, `title`, `answer`, `tags`, `similarity` (relevance score), `source`, `verified`) and `signal_id` for linking feedback. Note: `similarity` is an RRF score in hybrid mode (~0–0.02) or cosine similarity in semantic mode (0–1). The scale depends on the search mode — do not compare values across modes.
 
 ## Tips
 

skills/zeabur-rag-triage/SKILL.md

@@ -1,11 +1,11 @@
 ---
 name: zeabur-rag-triage
-description: List pending knowledge base maintenance items — open reports, unverified learned chunks, low-similarity queries, and negative feedback. Use when checking KB health, reviewing the maintenance queue, or starting a curation session. Triggers on "review queue", "what needs attention in the KB", "pending reports", "unverified chunks", "failed queries", "KB health check". Requires admin scope.
+description: List pending knowledge base maintenance items — open reports, unverified learned chunks, low-score queries, and negative feedback. Use when checking KB health, reviewing the maintenance queue, or starting a curation session. Triggers on "review queue", "what needs attention in the KB", "pending reports", "unverified chunks", "failed queries", "KB health check". Requires admin scope.
 ---
 
 # RAG — Triage
 
-Show what needs attention in the knowledge base. Returns 4 categories of pending items in a single call.
+Show what needs attention in the knowledge base. Returns 5 categories of pending items in a single call.
 
 Base URL: `$ZEABUR_RAG_URL`
 Auth: `Authorization: Bearer $RAG_API_KEY` — **admin scope required**
@@ -20,17 +20,18 @@ curl -s "$ZEABUR_RAG_URL/api/admin/triage?days=14&similarity_threshold=0.4&limit
 | Param | Default | Description |
 |-------|---------|-------------|
 | `days` | `14` | Time window for signals (reports and learned have no time window) |
-| `similarity_threshold` | `0.4` | Queries with `top_similarity` below this are flagged |
+| `similarity_threshold` | `0.4` | Queries with `top_similarity` below this are flagged. Note: in hybrid mode (RRF), scores are ~0–0.02 so the default 0.4 catches all hybrid queries. Adjust to ~0.005 for meaningful hybrid-mode filtering. |
 | `limit` | `20` | Max items per category |
 
 ## Response
 
-Returns JSON with 4 arrays + `counts` (true totals, not truncated) + `has_more` (boolean per category):
+Returns JSON with 5 arrays + `counts` (true totals, not truncated) + `has_more` (boolean per category):
 
 - **`open_reports`** — reports with `status='open'`, no time window. Fields: `id`, `type`, `chunk_id`, `query`, `detail`, `created_at`.
 - **`unverified_learned`** — learned chunks with `status='unverified'`, no time window. Fields: `id`, `title`, `tags`, `created_at`, `age_days`.
-- **`low_similarity_signals`** — queries where the best result scored below the threshold, within the time window, deduped by query text. Fields: `id`, `query`, `top_similarity`, `top_chunk_ids`, `client`, `created_at`.
+- **`low_similarity_signals`** — queries where the top score fell below the threshold, within the time window, deduped by query text. The score is RRF in hybrid mode (~0–0.02) or cosine in semantic mode (0–1). Fields: `id`, `query`, `top_similarity`, `top_chunk_ids`, `client`, `created_at`.
 - **`negative_feedback_signals`** — queries that got negative user feedback, within the time window. Fields: `id`, `query`, `feedback_score`, `feedback_comment`, `top_chunk_ids`, `created_at`.
+- **`needs_frontmatter`** — chunks whose tags include `"needs-frontmatter"`, meaning they were ingested from a source file that lacks proper frontmatter and got a temporary ID. Fields: `id`, `title`, `tags`, `url`, `source`, `created_at`.
 
 If a category query fails, its array is empty and a `<category>_error` field appears with the error message.
 

src/server.ts

@@ -1182,9 +1182,13 @@ const server = serve({
       "/": "ui/index.html",
       "/index.html": "ui/index.html",
       "/dashboard": "ui/dashboard.html",
-      "/report": "ui/report.html",
-      "/learn": "ui/learn.html",
     };
+    // Redirect old standalone pages to tabbed UI
+    if (url.pathname === "/report" || url.pathname === "/learn") {
+      const tab = url.pathname.slice(1);
+      const qs = url.search ? "&" + url.searchParams.toString() : "";
+      return Response.redirect(new URL("/?tab=" + tab + qs, url.origin), 302);
+    }
     if (url.pathname in uiRoutes) {
       if (!checkBasicAuth(req)) return unauthorizedResponse();
       return new Response(Bun.file(uiRoutes[url.pathname]), {