zeabur-rag-edit

RAG — Edit

Update the content of an existing chunk in place. Embedding is recomputed automatically when the searchable text changes, and the BM25 cache is cleared so the change is live on the next query.

Base URL: $ZEABUR_RAG_URL Auth: Authorization: Bearer $RAG_API_KEY — admin scope required

When to use

Situation	Use this skill
Fixing a typo or stale fact in an existing chunk	✅
Adjusting tags or toggling visibility (public ↔ internal)	✅
Rewriting the answer because the procedure changed	✅
You don't have admin scope	❌ Use zeabur-rag-report instead
The knowledge doesn't exist yet	❌ Use zeabur-rag-learn instead

API

PATCH /api/admin/chunks/:id — pass any subset of editable fields.

curl -s -X PATCH "$ZEABUR_RAG_URL/api/admin/chunks/<chunk_id>" \
  -H "Authorization: Bearer $RAG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "title": "Updated title",
    "answer": "Corrected content",
    "text_content": "Updated title\nCorrected content"
  }'

Field	Type	Notes
title	string	Display title
question	string	Question form (mainly for forum chunks)
answer	string	Main body
tags	string[]	Replaces the existing tag array
url	string	Source URL
text_content	string	The text that gets embedded and BM25-indexed
visibility	"public" | "internal"	Hard-filters search results
status	"unverified" | "verified" | "rejected"	Lifecycle state

Critical rule — keep the embedding in sync

If you update any of title / question / answer, you MUST also pass a new text_content. The server enforces this and returns 400 otherwise. The reason: embedding and BM25 index are built from text_content, not from the display fields — letting them drift apart silently corrupts search quality.

How to construct text_content:

Source	Convention
learned	${title}\n${answer}
forum (SUP-*)	Question: ${question}\n\nAnswer: ${answer}
docs (DOC-*)	Use the exact body text that was originally indexed; if unsure, fetch the chunk first via GET /api/admin/chunks/:id and adapt

When in doubt, read the existing text_content first, then rewrite it in the same shape.

Edits that don't need text_content

Pure metadata changes — tags, url, visibility, status — don't touch the embedding, so text_content is not required:

curl -s -X PATCH "$ZEABUR_RAG_URL/api/admin/chunks/SUP-1234" \
  -H "Authorization: Bearer $RAG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"tags": ["billing", "stripe"], "visibility": "internal"}'

Workflow

1. Find the chunk — search via zeabur-rag-search or GET /api/admin/chunks?search=... to get its ID and current contents.
2. Read it — GET /api/admin/chunks/:id returns the full record including text_content. Confirm you're editing what you think you're editing.
3. PATCH — send only the fields you want to change. Include text_content if any of title/question/answer are in the patch.
4. Verify — re-search to confirm the new content surfaces.

Response

{
  "success": true,
  "id": "SUP-1234",
  "updated_fields": ["title", "answer", "text_content"],
  "embedding_updated": true
}

embedding_updated: true means the chunk was re-embedded. Every edit is logged to rag_audit_log with action edit and a diff of old vs new values, and the BM25 cache is cleared so the next query sees the change.

Errors

Status	Cause
400	Updated title/question/answer without providing text_content, or invalid visibility/status/tags
401 / 403	Missing/invalid key, or key lacks admin scope
404	Chunk ID does not exist
500	DB or embedding service error — safe to retry

Guidelines

• One coherent edit at a time. Don't rewrite three unrelated chunks in one batch — each edit should be auditable.
• Don't paraphrase for style. Edit only when the content is wrong, stale, or unclear. Cosmetic rewrites churn the embedding for no quality gain.
• Verified chunks should stay accurate. If you're unsure whether your fix is correct, file a zeabur-rag-report instead of patching.
• Visibility changes are powerful. Flipping public → internal immediately hides the chunk from public search. Confirm before doing this.