Knowledge Base Sync

How sync works

TR-AI crawls your KB sources and generates vector embeddings for each content chunk. These embeddings are stored in the database and used for similarity search when a ticket arrives.

The sync process:

• Crawls your website starting from the root URL, following internal links
• Fetches pages from Confluence or Notion via API
• Splits content into ~400-token chunks
• Generates 384-dimension embeddings using the all-MiniLM-L6-v2 model (local WASM)
• Stores chunks and embeddings in the database
• Uses SHA-1 deduplication to skip unchanged pages

Automatic sync by plan

Under Settings → Knowledge base sync you choose how often TR-AI checks your documentation for changes (it only re-embeds pages that changed). Each plan sets the most frequent interval; choosing a slower one is always possible.

After a downgrade, a previously chosen faster interval is kept but auto-sync runs at the new plan's interval.

Triggering a manual re-index

Trigger a re-index from the Dashboard → Knowledge Base card → Re-index button. This rebuilds all embeddings from scratch.

You can also trigger a re-index via the API:

POST /api/admin/reindex
Content-Type: application/json

// Re-index a specific trial account:
{ "trialAccountId": 42 }

// Re-index all accounts (admin only):
{}

Auto re-index on startup

On server startup, TR-AI automatically detects trial accounts that have KB chunks but no 384-dimension embeddings (e.g. accounts that were indexed with the old 1536-dimension OpenAI model). These accounts are re-indexed in the background automatically.

While a re-index is running, tickets for that account are routed directly to the agent (routed_to_agent outcome) to avoid low-quality resolutions.

Vector stats

Check the current state of your embeddings via the API:

GET /api/admin/vector-stats?trialAccountId=42

Response:
{
  "chunks": [
    { "model": "minilm384", "dims": 384, "count": 1247 },
    { "model": "openai",    "dims": 1536, "count": 0 }
  ]
}

A healthy account should show a non-zero count for minilm384 (384 dims) and zero for openai (1536 dims). If you see 1536-dim vectors, trigger a re-index to rebuild with the local model.

KB source limits