Collections — Data Templates

Collections are the structure of your organization's database in Rela AI. They define what information you store, what fields it has, and how it is organized. Think of them as the "sheets" of a database: "Assets," "Maintenance Invoices," "Equipment Technical Sheets," "Suppliers," "Spare Parts Inventory."

A well-defined collection is the foundation that allows AI agents to answer precise questions about your plant's data.

What is it for?

Industrial plant data exists in multiple forms: manuals in PDF, paper invoices, technical sheets in Excel, specifications in Word. Without a clear structure, AI agents cannot query that information — it would be like asking someone to find a document in a disorganized archive with no index.

Collections solve this by defining:

• What fields each type of information has (e.g., an equipment sheet has manufacturer, model, serial number, operating parameters)
• What data type each field is (text, number, date, file)
• What each field means (description to guide the AI)

Once the structure is defined, you can populate the collection with records — either manually or through automatic document extraction.

How does it work?

Each record stored in a collection automatically generates an embedding — a mathematical representation of the content's meaning. This enables semantic search: when the agent searches for "hydraulic failure air compressor," it can find a record that says "pneumatic system fault central compressor," even though the exact words do not match.

How to use it?

Create a collection

1. Go to Data > Extractions in the sidebar.
2. Click New Collection.
3. Define a descriptive name for the collection (e.g., "Equipment Technical Sheets").
4. Add a description explaining what type of information it contains.
5. Define the fields:

6. Save the collection.

Available field types

Typical collections for industrial maintenance

"Equipment Technical Sheets":

• manufacturer (string): "Siemens", "Grundfos", "Atlas Copco"
• model (string): "1LA7 096-4AA10", "CR 32-4"
• serial_number (string): "SIE-2023-45678"
• power_kw (float): 7.5
• nominal_voltage (int): 380
• max_temperature (float): 85.5
• recommended_lubricant (string): "Shell Omala S2 G 220"
• manual_pdf (file): PDF file of the manual

"Spare Parts Inventory":

• part_code (string): "SKF-6205-2Z"
• description (string): "SKF deep groove ball bearing"
• current_stock (int): 8
• minimum_stock (int): 3
• supplier (string): "SKF North America"
• unit_price (float): 245.50
• warehouse_location (string): "Shelf A-12"

"Maintenance Invoices":

• invoice_number (string): "INV-2026-00234"
• supplier (string): "TechMaint Inc."
• issue_date (date): "2026-03-15"
• work_description (string): "Bearing replacement motor M-12"
• affected_equipment (string): "Motor M-12"
• total (float): 4850.00

Edit a collection

You can update the name, description, and add new fields to an existing collection. Changes to the structure do not affect already-saved records — they only apply to new records.

Delete a collection

Deleting a collection removes its definition and structure. Existing records remain in the database but lose their link to the template.

Connection with the rest of the system

Collections are used in three main flows:

1. Extractions: when you upload a PDF, the AI extracts the fields defined in the collection
2. Records: extracted or manually entered data is saved as collection records
3. Agent tools: WhatsApp and email agents can query collections to answer questions

Key benefits

• Flexible structure that adapts to any type of plant information
• Automatic semantic search — agents find information by meaning
• Foundation for automatic document extraction with AI
• File-type fields to store PDFs and images linked to each record
• Multiple collections to organize different types of information

Common use cases

Scenario 1: Equipment knowledge base The maintenance coordinator creates the "Technical Sheets" collection with fields for all technical data on the plant's equipment (manufacturer, model, parameters, lubricants, bearings). They load the sheets for 45 critical assets, half manually and half by extracting data from technical manuals in PDF. Technicians can now ask the WhatsApp agent any technical specification without searching through physical files.

Scenario 2: Queryable inventory control The warehouse maintains a "Spare Parts Inventory" collection with part code, description, current stock, and minimum stock. When the stock of a part reaches the minimum, the agent detects the record (which has a field "low_stock_alert" = true) and automatically notifies the buyer.

Scenario 3: Supplier history Procurement maintains a "Maintenance Suppliers" collection with service history, response times, prices, and quality ratings. The coordinator can ask the agent "who is the most reliable supplier for alignment work?" and the agent searches the collection and responds with the top-rated suppliers for that type of job.