AI Assistant Overview

The AI Assistant is your conversational partner for building and managing web automations. Instead of manually configuring commands, selectors, and datasets, you simply describe what you want in plain English, and the AI builds it for you.

The AI Assistant can analyze web pages, generate project configurations, create datasets, set up schedulers, debug failed runs, and answer questions about your projects — all through a simple chat interface.

Changed in v8.3.2: The AI Assistant now uses a full-context injection strategy instead of RAG (Retrieval-Augmented Generation). The complete system prompt and command reference are provided to the AI in every conversation, ensuring it always has access to every command, parameter, and strategy. See Release Notes for details.

AI Assistant chat interface showing a conversation where the user asks to scrape a website and the AI generates a project configuration

What the AI Assistant Can Do

The AI Assistant is a versatile tool that can help you with many aspects of web automation:

Capability	Description
Build automation projects	Describe what you want to automate, and the AI generates the complete project configuration.
Analyze web pages	The AI opens the target page, examines its structure, and identifies the best selectors and data patterns.
Generate datasets	The AI identifies repeating data patterns on a page and creates dataset configurations automatically.
Create schedulers	Ask the AI to schedule a project to run at specific intervals.
Debug failed runs	Share error logs with the AI, and it will suggest fixes for broken selectors or commands.
Answer questions	Ask about RTILA X features, commands, or best practices.
Modify existing projects	Chat about a specific project and ask the AI to make changes.

AI Providers

RTILA X supports multiple AI providers, giving you flexibility in how you power the AI Assistant:

Provider	Description	Best For
OpenRouter	Access to hundreds of cloud AI models (DeepSeek, Claude, Gemini, GPT, and more) through a single API key.	Users who want the best quality and don’t mind cloud processing.
Local LLM (GGUF)	Download and run any GGUF model from Hugging Face directly on your computer using the bundled llama.cpp server.	Maximum privacy. No internet needed after download.
Local LLM (Custom API)	Connect to any OpenAI-compatible API endpoint (Ollama, LM Studio, vLLM, text-generation-webui).	Users who already have a local AI server running.
RTILA Courtesy AI (Preview)	A complimentary shared AI provided by RTILA. No API key needed.	Getting started quickly. May change over time.

Choosing a Provider

For best quality: Use OpenRouter with a model like google/gemini-3.5-flash-lite or anthropic/claude-sonnet-4.5.
For maximum privacy: Use Local LLM (GGUF) with a model like Qwen 2.5 7B or Llama 3.1 8B.
For quick testing: Use the RTILA Courtesy AI (no setup required).

Tip: You can switch providers at any time in Preferences → AI Assistant. Your chat history is preserved regardless of which provider you use.

Context Strategy: Full Injection (Not RAG)

In v8.3.2, RTILA X uses a full-context injection strategy for the AI Assistant. This means:

The complete system prompt (behavioral instructions for the AI) is included in every request.
The complete command reference (every command, parameter, selector strategy, and configuration rule) is included in every request.
The AI always has access to the full knowledge base — no retrieval, no search, no missing context.

Why Not RAG?

Previous versions used RAG (Retrieval-Augmented Generation), which searched a vector database for relevant context. This was removed because:

Retrieved chunks were often irrelevant or misleading.
RAG added 350–850ms of latency per message.
The embedding model lacked domain specificity for RTILA commands.
Full-context injection with KV caching is faster and more reliable.

How It Works Now

Provider	Context Strategy
OpenRouter	Full system prompt + complete reference injected into every request.
Local LLM (GGUF)	Full system prompt + complete reference injected once per KV cache slot, reused across all subsequent messages.
Local LLM (Custom API)	Full system prompt + complete reference injected into every request.

For local LLM users, the KV cache means the system prompt and reference are processed once and then reused for every subsequent message in the conversation. This makes the full-context approach effectively free after the first message.

Chat Settings

Each chat conversation has configurable settings. Click the gear icon next to the chat title to open the Chat Settings dialog.

Chat Settings dialog showing title input, project examples count, and message history dropdown

Settings Table

Setting	Type	Default	Description
Chat Title	Text	First 50 characters of first message	A display name for the chat. Helps you identify conversations in the sidebar.
Project Examples in Context	Number (0–49)	0	How many recent projects to include as examples in the AI’s context. Helps the AI understand your project structure and style.
Message History	Dropdown	Last 7 messages	How many past messages to include in the AI’s context for each new request.

Message History Options

Option	Value	Best For
Last 7 messages	7	Short conversations. Uses less context.
Last 15 messages	15	Medium conversations. Good balance.
Last 30 messages	30	Long conversations. More context for the AI.
Unlimited	-1	Very long conversations. Uses the most context.

Tip: If you’re using a local LLM with limited context (e.g., 8K tokens), keep the message history at 7 to avoid running out of context. For cloud models with 128K+ context, you can safely use Unlimited.

Project Examples in Context

When you set “Project Examples in Context” to a number greater than 0, the AI includes that many recent projects as examples in its context. This helps the AI:

Understand your naming conventions.
Match your project structure and style.
Generate configurations that are consistent with your existing projects.

Warning: Setting this too high (e.g., 49) uses a lot of context. If you’re using a local LLM, keep this at 0–5 to avoid context overflow.

AI Tools

The AI Assistant has access to several tools that it can use to analyze web pages and interact with your browser:

Tool	Description
analyze_target_page	Opens the target URL in a stealth browser, analyzes the page structure, identifies elements, forms, data patterns, and network requests, and returns a detailed report.
navigate_and_analyze	Navigates to a secondary page (e.g., a product detail page or a pagination page) for additional analysis.
request_user_input	Prompts you for credentials, CAPTCHA solutions, or other information when the AI encounters a login wall or protected content.

How Tools Work

When you ask the AI to build a project for a specific URL, it:

Calls analyze_target_page with the URL.
A stealth browser opens the page (using your stealth settings).
The AI examines the page structure, identifies elements, and captures a screenshot.
The analysis results are returned to the AI.
The AI uses this information to generate the best commands and selectors.

Tool Call Indicators

When the AI is using a tool, you’ll see an indicator in the chat:

“AI is launching a stealth browser to analyze [URL]…”

And when the analysis completes:

“Page Analysis Complete (Found X elements, Y forms)”

Multi-Page Analysis Workflow

For complex scraping tasks, the AI may need to analyze multiple pages. Here’s how the multi-page workflow works:

You describe the task. For example: “Scrape product details from this listing page, including the detail pages.”
The AI analyzes the listing page. It calls analyze_target_page on the main URL.
The AI identifies detail page links. It recognizes that product cards link to detail pages.
The AI analyzes a detail page. It calls navigate_and_analyze on a sample detail page URL.
The AI generates the complete configuration. It creates commands for navigating to the listing, extracting links, visiting each detail page, and extracting the data.

This multi-page analysis ensures the AI understands the full data structure before generating the project.

AI Self-Correction

When AI Self-Correction is enabled in your project settings, the AI can automatically fix errors during project execution.

How It Works

During execution, if a run_script command fails, the error is captured.
The error message and the failed script are sent to the AI.
The AI analyzes the error and generates a corrected version of the script.
The corrected script replaces the failed one.
The command is retried.

Retry Limits

The AI will attempt self-correction up to 3 times per command.
After 3 failed attempts, the command fails permanently and the project execution stops.
You’ll see a log message indicating the self-correction attempt: “Self-Correction successful. Retrying the automation command (Attempt X/3)”

Enabling Self-Correction

Select a project.
Click Settings in the ribbon.
Go to the Advanced tab.
Check Enable AI Self-Correction.
Click Save.

Tip: Self-correction is especially useful for websites that change their HTML structure frequently. The AI can adapt to minor selector changes automatically.

Handling Protected Pages

When the AI encounters a protected page (login wall, CAPTCHA, 2FA), it uses the request_user_input tool to ask you for help.

What Happens

The AI detects a login form, CAPTCHA, or 2FA prompt on the page.
It calls request_user_input with a description of what’s needed.
A modal dialog appears in RTILA X asking for the required information.
You enter your credentials or solve the CAPTCHA.
The AI uses your input to proceed with the analysis.

Example Prompts

Protection Type	What the AI Asks
Login form	“I found a login form. Please provide the username and password fields.”
CAPTCHA	“I encountered a CAPTCHA. Please solve it or provide the solution.”
2FA	“The site requires two-factor authentication. Please provide the code.”
Cookie consent	“I found a cookie consent banner. Should I accept or reject it?”

Note: Your credentials are used only for the current analysis session. They are not stored permanently unless you add them as project variables.

Chat Context Sources

The AI builds its context from multiple sources. Understanding these helps you get the best results:

Context Source	Description
System prompt	Base behavioral instructions telling the AI how to act as an RTILA automation architect.
Complete command reference	Every command, parameter, selector strategy, and configuration rule. Always included.
Project examples	Recent project configurations (if configured in chat settings).
Current project	The project associated with the chat (if any).
Message history	Previous messages in the conversation (limited by the Message History setting).
Tool results	Results from `analyze_target_page` and `navigate_and_analyze` calls.

Changed in v8.3.2: The previous RAG knowledge base context has been replaced by the complete command reference, which is always available to the AI without retrieval errors or relevance tuning.

Chatting About a Specific Project

You can start a chat that’s linked to a specific project. This gives the AI full context about that project, so you can ask it to make changes or debug issues.

Steps

Go to Automation Projects.
Select the project you want to chat about.
Click Chat AI in the ribbon toolbar.
A new chat opens, linked to that project.
The AI now has full context about the project’s commands, datasets, and settings.

What You Can Ask

Once you’re chatting about a specific project, you can ask things like:

“Change the selector for the product title to use the data-product-name attribute.”
“Add a command to wait for the page to fully load before extracting data.”
“Why did my last run fail? Here are the logs.”
“Add a new dataset to extract the product images.”

Keyboard Shortcuts

The AI Assistant chat input supports the following keyboard shortcuts:

Shortcut	Action
Enter	Send the message.
Shift+Enter	Insert a new line without sending.

Tip: Use Shift+Enter to write multi-line prompts. This is useful for complex requests with multiple instructions.

The Chat Interface

The AI Assistant chat interface consists of:

Element	Description
Chat sidebar	Shows all your chat conversations. Click to switch between them.
Chat messages	The conversation history. User messages appear on the right, AI messages on the left.
Chat input	The text box at the bottom where you type your messages.
Attachment button	Click the + button to attach files or access additional actions.
Send button	Click to send your message.
Stop button	Appears while the AI is generating. Click to stop generation.

Message Actions

Each AI message has a copy button that appears when you hover over it. Click it to copy the message to your clipboard.

Running Projects from Chat

When the AI generates a project configuration, you’ll see a Run Project button below the configuration. Click it to immediately run the generated project.

Viewing Results in Chat

After a project run completes, you can ask the AI to fetch and analyze the results. The AI will display the data in a formatted table within the chat.

Tips and Best Practices

Tip	Why It Matters
Be specific about what you want to extract.	“Extract product names, prices, and image URLs” is better than “get the data.”
Include the full URL.	The AI needs the exact URL to analyze the page.
Mention if the site requires login.	The AI will prompt you for credentials.
Use the “Project Examples in Context” setting.	Helps the AI match your project style.
Keep message history reasonable.	Too much history uses too much context.
Use Shift+Enter for multi-line prompts.	Helps you write complex, detailed instructions.
Check the tool indicators.	They tell you when the AI is analyzing a page.
Enable Self-Correction for dynamic sites.	The AI can fix broken selectors automatically.
Chat about specific projects for modifications.	The AI has full context and can make targeted changes.
Use the RTILA Courtesy AI for quick tests.	No API key needed. Add your own key for guaranteed access.