Key Concepts
Core terminology and fundamental concepts of the Rela AI platform.
Agent
An agent is a conversational AI configured to interact with users through a specific channel. Rela AI supports two types of agents:
| Channel | AI Engine | Integration |
|---|---|---|
| AI engine | Evolution API (WhatsApp Business) | |
| AI engine | Postmark (email delivery) |
Each agent has:
- Name — Unique identifier.
- System prompt — Instructions organized into editable sections that define the personality and behavior.
- Tools — Capabilities the agent uses during the conversation (search data, generate reports, call APIs).
- Configuration — Channel-specific parameters (AI model, language, timeouts).
Tool
A tool gives your agent capabilities beyond just conversing. There are three types:
- Data query — Searches for information in your collections using semantic search (by meaning) or exact search (by code/serial number). You configure the collection, search fields, and maximum results.
- Internal action — Executes predefined system functions:
generate_report(PDF),assign_task(tasks),send_whatsapp_message,send_email,locate_nearest_personnel,send_file. Each action accepts parameters with optional default values. - External connection — Connects to systems outside the platform using HTTP (REST APIs), MQTT (lightweight messaging protocol for IoT devices), or OPC UA (standard protocol for communication with industrial machinery) such as Siemens, Beckhoff, and Allen-Bradley PLCs.
The agent automatically decides when to use each tool based on the conversation and the tool's description.
Collection
A collection is a set of structured data with typed fields (string, float, int, boolean, date, list, file). They can be created in two ways:
- By extraction — AI extracts data from documents (PDFs, images) automatically according to a field template.
- Manually — You define the fields and add records directly.
Each record automatically generates an embedding (a numerical representation that enables smart search) that powers semantic search. For example, searching for "laptop for design" finds records like "MacBook Pro for graphic design" even though the words don't match exactly.
Collections also support exact search for fields like codes, serial numbers, or IDs, where you need precise matching.
System Prompt
The system prompt defines how your agent behaves. It is organized into independent sections that you can edit, reorder, and optimize individually:
| Section | Purpose |
|---|---|
| Identity | Who the agent is — name, role, personality, tone |
| Instructions | What it should do — conversation flow, steps, information to collect |
| Rules | Limits and restrictions — what it should NOT do, escalation policies |
| Tools | How to use the tools — when to invoke them, what parameters to request, how to present results |
Each section has an Enhance with AI button that optimizes the content for the specific channel (WhatsApp: formatted text, email: professional text). Additionally, you can add custom sections for FAQs, schedules, special procedures, etc.
Extraction
A process where AI analyzes documents and extracts structured data. It uses computer vision for scanned PDFs and images, and text processing for documents with selectable text. It supports:
- Individual extraction — One record per document.
- Batch extraction — Multiple records from a single document (tables, catalogs, inventories).
- Audit — Generates annotated images showing where each piece of data was found in the original document.
Tenant
Rela AI is multi-tenant. Each organization (tenant) has complete data isolation: its own agents, collections, tools, departments, staff, conversation history, and reports. No tenant can access another tenant's data.