Vector Store and RAG Integration

This is the technical reference for integrating your custom agents with Agent Builder’s Vector Store and RAG (Retrieval-Augmented Generation) pipeline. For the user-facing overview, see AI Data Training.

Architecture Overview

Agent Builder’s RAG system has three layers:

1. Ingestion — content is chunked, embedded, and stored in a namespace-isolated vector index on Google Vertex AI
2. Retrieval — at query time, the user’s message is embedded and matched against stored vectors using semantic similarity search
3. Generation — retrieved passages are injected into the LLM context as grounding evidence, producing answers that cite your actual documents

All communication goes through the Agentic RAG API at https://rag.agentic-plugin.com. Authentication uses a per-site API secret stored in agentic_rag_api_secret.

Training Sources

There are two ways to feed content into the vector store:

1. Train on WordPress Content

From Agent Builder → Train on Data, the admin UI scans your published posts, pages, and custom post types. Select which content to train and click Train. Behind the scenes, the plugin calls:

POST /train/text
{
  "user_id": "site-unique-id",
  "source_id": "post-123",
  "title": "Post Title",
  "content": "The full post content as plain text...",
  "url": "https://yoursite.com/post-slug/",
  "metadata": {
    "post_type": "post",
    "author": "admin"
  }
}

Content is chunked server-side into semantically meaningful segments (typically 500–1,000 tokens) with overlap, then embedded using Google’s text-embedding model and indexed in your namespace.

2. Upload Files (PDF, TXT)

Upload documents directly from the Train on Data page. The plugin sends a multipart upload to:

POST /train (multipart/form-data)
Fields: user_id, file

Supported formats: PDF, plain text. Files up to 50 MB. The API extracts text, chunks, embeds, and indexes in the same pipeline as WordPress content.

RAG API Endpoints

All endpoints require the X-API-Key header with your site’s RAG secret.

Method	Endpoint	Purpose
POST	/train/text	Train on text content (posts, pages)
POST	/train	Upload and train on a file (multipart)
GET	/sources?user_id=X	List all trained sources
DELETE	/sources	Delete a specific source by ID
GET	/query?q=X&user_id=X	Semantic search across trained content

How Agents Use RAG Context

When a user sends a message to an agent on a site with trained data, the chat pipeline automatically:

1. Queries the vector store — the user’s message is sent to /query to find the most relevant passages from your trained content
2. Injects context — matching passages are prepended to the system prompt as grounding evidence, with source attribution
3. Generates response — the LLM receives both the user’s question and the retrieved passages, producing an answer grounded in your actual data
4. Cites sources — the response includes references to the source document and passage, so users can verify the answer

This happens transparently — your agent code does not need to call RAG tools explicitly. Any agent on a site with trained data automatically gets RAG-augmented context.

Programmatic Access

Developers can interact with the RAG system programmatically using the RAG_Manager class:

// Query the vector store
$results = \Agentic\RAG_Manager::api_request( '/query', 'GET', [
    'q'       => 'What is our refund policy?',
    'user_id' => get_option( 'agentic_site_id' ),
    'top_k'   => 5,
] );

// Train on custom content
$result = \Agentic\RAG_Manager::api_request( '/train/text', 'POST', [
    'user_id'   => get_option( 'agentic_site_id' ),
    'source_id' => 'custom-doc-1',
    'title'     => 'Refund Policy',
    'content'   => $my_document_text,
    'url'       => 'https://yoursite.com/refund-policy/',
] );

// List all trained sources
$sources = \Agentic\RAG_Manager::api_request( '/sources', 'GET', [
    'user_id' => get_option( 'agentic_site_id' ),
] );

// Delete a source
$deleted = \Agentic\RAG_Manager::api_request( '/sources', 'DELETE', [
    'user_id'   => get_option( 'agentic_site_id' ),
    'source_id' => 'custom-doc-1',
] );

AJAX Endpoints

The Train on Data admin page uses these WordPress AJAX actions (all require the agentic_train_data nonce):

AJAX Action	Purpose
agentic_td_get_overview	Get training stats and WordPress content scan
agentic_td_scan_content	Scan published posts/pages for training candidates
agentic_td_train_post	Train a single post by ID
agentic_td_upload_file	Upload and train on a document file
agentic_td_get_sources	List all trained sources from the vector store
agentic_td_delete_source	Remove a source from the vector store
agentic_td_get_credits	Check remaining credit balance
agentic_td_get_pricing	Fetch current credit pricing
agentic_td_get_transactions	Get credit transaction history

Data Isolation and Security

• Namespace isolation — every site gets its own vector namespace. Cross-tenant retrieval is impossible.
• Encryption — TLS 1.3 in transit, AES-256 at rest. Keys managed by Google Cloud KMS.
• No model training — your data is never used to train Google’s foundation models. Governed by Google’s Data Processing Addendum.
• Right to erasure — delete any source or your entire corpus at any time, effective immediately.

Credit Costs

Operation	Cost	Unit
Embed / Train	1 credit	per 1,000 tokens (~750 words)
Semantic Query	1 credit	per search query
Delete	Free	always

Credits are shared across all Agentic services (RAG, Image Generation, Text-to-Speech). Check your balance in Settings → Health → Credit Balance.

Requirements

• Active Personal or Agency license
• Prepaid credits (check balance in Settings → Health)
• PHP 8.1+ with cURL extension

Related

• AI Data Training — user-facing overview and enterprise use cases
• Building a Custom Agent — agent code structure
• Agent Tools — all available built-in tools
• GDPR and Data Protection — privacy compliance