Skip to main content

Admin API Endpoints

All admin endpoints require verifyAdmin() authentication — the user must have a valid JWT and their email must exist in the admin_users table.

RAG Content Management​

POST /api/admin/submit-content​

Upload new content to the RAG knowledge base (v2 architecture: Supabase is the source of truth, Pinecone holds vectors).

{
  "uploadType": "html",
  "htmlContent": "<!DOCTYPE html>...",
  "description": "Product showcase with tabbed navigation",
  "technique": "tabs",
  "complexity": "intermediate",
  "emailPurpose": "ecommerce",
  "keyFeatures": ["lightswitch", "mobileResponsive"],
  "bestPracticeTags": ["tableStructure", "msoConditionals"]
}
FieldTypeRequiredDescription
uploadTypestringYeshtml, blog, or amp
htmlContentstringYes (html/amp)The email HTML
descriptionstringYes (html/amp)Content description
techniquestringNoKinetic technique (tabs, carousel, etc.)
complexitystringNobeginner, intermediate, advanced
emailPurposestringNoe.g. ecommerce, newsletter, transactional
exampleTypestringNopositive (default) or negative example
keyFeatures / bestPracticeTagsstring[]NoTags for categorization
embeddingTextstringNoHuman-reviewed embedding text (overrides the auto-built text)

The endpoint:

  1. Builds embedding text (technique-specific semantic descriptions injected), or uses the provided embeddingText
  2. Generates a single embedding with OpenAI text-embedding-3-large (3072 dims)
  3. Stores the full document — HTML, metadata, and embedding text — in the Supabase rag_documents table (service-role only, no size cap)
  4. Upserts the vector plus small metadata into the Pinecone v2 index

POST /api/admin/update-content​

Update existing RAG content: { "id": "...", "metadata": { ...updated fields } }. Rebuilds the embedding text and re-upserts with fresh embeddings.

POST /api/admin/delete-content​

{
  "id": "content-uuid"
}

GET /api/admin/list-content​

Browse the full RAG library. Lists every vector ID via Pinecone listPaginated (no topK cap — the old dummy-vector query silently dropped data past 10,000 items), then batch-fetches metadata.

{
  "success": true,
  "count": 156,
  "items": [{ "id": "...", "metadata": { ... } }]
}

Items are sorted by submission date, newest first.

GET /api/admin/rag-stats​

Returns statistics about the RAG knowledge base:

{
  "totalDocs": 156,
  "totalChars": 2450000,
  "htmlDocs": 98,
  "blogDocs": 42,
  "ampDocs": 16,
  "avgCharsPerDoc": 15700,
  "documents": [{ "id": "...", "type": "html", "chars": 12400 }]
}

Eval Harness​

GET/POST /api/admin/eval​

Service-role storage for the eval harness — the eval tables stay locked behind the server (no client RLS).

CallAction
GETList recent runs (up to 50)
GET ?run=<id>One run plus its results
POST { "action": "createRun", "label": "...", "promptSetVersion": "...", "promptCount": 12 }Create a run, returns { runId }
POST { "action": "saveResult", "runId": "...", "result": { ... } }Store one prompt's result (QA score, judge score, gen time, error)
POST { "action": "finalizeRun", "runId": "..." }Compute averages and mark complete. Errored prompts count as 0 in averages

POST /api/eval/judge​

LLM judge (Claude Haiku) scores a generated email for the eval harness. Admin-only.

Utilities​

POST /api/admin/auto-tag​

Uses Claude to automatically generate metadata tags for content:

{
  "htmlContent": "<!DOCTYPE html>...",
  "contentType": "html"
}

Returns:

{
  "description": "Tabbed product showcase with radio button navigation and lightswitch fallback",
  "technique": "tabs",
  "complexity": "intermediate",
  "emailPurpose": "promotional",
  "keyFeatures": ["lightswitch", "mobileResponsive"],
  "bestPracticeTags": ["tableStructure", "msoConditionals"],
  "embedding_text": "Dense structural + intent paragraph for retrieval..."
}

POST /api/admin/submit-feedback​

Collect feedback on generated emails:

{
  "prompt": "Original generation prompt",
  "rating": "positive",
  "ragUsed": true,
  "ragExamplesCount": 5
}