Retraining and Knowledge Sync

Whenever you add a new website URL, upload an updated product catalog, or adjust a text snippet, Pitaj.io automatically triggers an asynchronous background ingestion pipeline.

In this guide, you will learn how to monitor training states, initiate immediate syncs, and inspect retrieved context.

---

The Knowledge Ingestion Lifecycle

Each data source in your dashboard moves through the following processing states:

State	Definition	System Action
Queued	Source received in the processing backlog	A background worker is assigned to fetch and validate the resource
Processing	File or page is actively being parsed	Content is stripped of boilerplate, chunked semantically, and sent to embedding models
Trained	Vectors written to the index	The agent begins citing this information across all active sessions
Failed	An error occurred during ingestion	A descriptive error log is provided (e.g., password-protected PDF or HTTP 403 Forbidden)

---

Triggering a Manual Sync

If you have just pushed critical pricing or policy adjustments to your live site:

1. Navigate to Knowledge Base → Websites (or Files).
2. Locate the resource you want to refresh.
3. Click the three dots menu icon (...) on the right side of the row.
4. Select Re-train Now.
5. The pipeline retrieves the fresh data, updates embeddings, and replaces stale segments without causing downtime to visitor chats.

---

Character Quotas and Source Monitoring

At the bottom of the Knowledge Base section, you can monitor global usage indicators:

• Total Character Usage: Sum of all extracted alphanumeric characters across websites, PDFs, snippets, and Q&A items.
• Active Data Sources: Number of indexed links and uploaded files against your tier limit.

Storage Optimization

If you approach the character threshold of your plan, delete legacy catalogs or add exclusion filters to avoid indexing marketing blog posts that do not directly assist customer support.

---

Verifying Answers in Real-Time

We recommend testing new knowledge updates immediately inside the Live Preview pane on the right:

Source Attribution Inspection

Inside the test conversation pane, click the source citation icon beneath any response to inspect the exact textual chunk retrieved by RAG.

Similarity Score Threshold

If the agent seems overly hesitant or overly permissive, adjust the minimum semantic cosine similarity threshold in Advanced Model Settings.