Release Notes

This page documents all changes, new features, removed features, and bug fixes for each version of RTILA X.

Version 8.3.2 (Current)

Breaking Changes

RAG / Qdrant Vector Database — Removed

What was removed: The local RAG (Retrieval-Augmented Generation) pipeline backed by Qdrant has been completely removed from RTILA X.

Why it was removed:

After extensive testing across hundreds of real-world automation tasks, the RTILA team determined that RAG was hurting performance rather than helping it:

1. Reduced Accuracy: The embedding model (all-MiniLM-L6-v2) lacked the domain specificity needed for accurate retrieval of RTILA X command schemas. Retrieved chunks were frequently irrelevant to the user’s actual query, from outdated or deprecated command examples, or missing critical context (e.g., retrieving a command description without its parameter list).

2. Context Pollution: Injected RAG chunks displaced higher-value context in the AI’s limited context window. The system prompt, project configuration, and user instructions were truncated to make room for retrieved examples that were often misleading.

3. Latency Overhead: Each message required ~200ms for embedding generation, ~100–500ms for Qdrant vector search, and ~50ms for result formatting. This added 350–850ms of latency per message with no quality improvement.

4. Resource Consumption: The Qdrant binary consumed 400–800 MB RAM at idle, the embedding model required 1–2 GB RAM, Qdrant storage used 200–500 MB disk, and the Python venv for ingestion used 500 MB+ disk. Total: 1.5–3 GB RAM, 700 MB+ disk.

5. Full-Context Approach Is Superior: By injecting the complete system prompt and command reference directly into every conversation, the AI always has access to every command, parameter, and strategy simultaneously. No retrieval errors, no relevance threshold tuning, no embedding quality issues.

What replaces it:

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

Removed components:

• Qdrant binary (rtila-vector)
• @qdrant/js-client-rest npm dependency
• @xenova/transformers npm dependency
• RAG settings (Top K, Similarity Threshold, server status)
• initialize_qdrant, stop_qdrant, get_qdrant_status Tauri commands
• vector/storage/ bundled resource
• ingest_data.py ingestion pipeline
• Qdrant Python venv

Finetuned Models — Removed

What was removed: Custom finetuned models (RTILA Lite 1.5 Assistant and all prior rtila-corporation/* models) have been removed.

Why they were removed:

1. Maintenance Burden: Each model update required retraining, re-evaluation, re-quantization, re-upload, and regression testing across multiple hardware configurations. This cycle took 2–4 weeks per update.

2. Hallucination Issues: Finetuned models, particularly at Q4 quantization, exhibited significantly higher hallucination rates than base models. They would confidently generate invalid command names, invent non-existent parameters, and fabricate selector syntax not present in the training data.

3. Context Limitation: The 9B finetuned model had an effective context window of 8K–16K tokens, insufficient for the full command reference (which requires 32K+). This forced truncation of the knowledge base, defeating the purpose.

4. No Meaningful Quality Gain: In blind A/B testing, the finetuned model produced configurations of equal or lower quality compared to modern open-source base models (Llama 3.1 8B, Qwen 2.5 7B, Mistral 7B) when given the same system prompt and reference context.

5. Hardware Incompatibility: The finetuned models were optimized for specific hardware profiles. Users with different GPU/CPU combinations experienced crashes, OOM errors, or degraded performance.

What replaces them: The Local LLM provider now supports any GGUF model from Hugging Face in the user/repo/file.gguf format, plus any OpenAI-compatible API endpoint (Ollama, LM Studio, vLLM, text-generation-webui, etc.).

New Features

Local LLM Provider — Complete Redesign

The Local LLM provider has been completely redesigned with two modes:

Mode	Description
GGUF Model (Hugging Face)	Download and run any GGUF model via the bundled llama.cpp server. Format: user/repo/file.gguf. Downloaded automatically on first use.
Custom API Endpoint	Connect to any OpenAI-compatible API (Ollama, LM Studio, vLLM, etc.). No model download needed.

KV Cache Configuration

Full control over llama.cpp’s KV cache for optimal local performance:

Setting	Options	Default	Description
KV Cache Type K	F16, Q8_0, Q4_0, F32	F16	Quantization for the Key cache
KV Cache Type V	F16, Q8_0, Q4_0, F32	F16	Quantization for the Value cache
Flash Attention	On / Off	Off	Enables flash attention kernels
Parallel Slots	1–8	1	Concurrent conversation slots
No MMap	On / Off	Off	Disables memory-mapped loading
MLock	On / Off	Off	Locks model weights in RAM

Token Optimization

• System prompt + complete reference injected once per slot (KV cached)
• Context truncation guard prevents overflow on smaller models
• Default context window increased to 32K for local models
• Estimated token count displayed in chat settings

Enhanced Error Diagnostics

When the local LLM fails to start, RTILA X now provides hardware-specific guidance:

Error Type	What You’ll See	Recommended Action
Out of Memory	“Your hardware lacks sufficient RAM/VRAM for the requested context size”	Reduce Context Size or GPU Layers, or switch to OpenRouter
GPU Driver Error	“GPU driver may be outdated or incompatible”	Update GPU drivers, or set GPU Layers to 0
Model Load Error	“GGUF file may be corrupted or incompatible”	Re-download the model

All errors include: “This is a hardware compatibility issue, not a software bug. You can switch to OpenRouter or a Custom API endpoint in Preferences > AI Assistant to use cloud-based models instead.”

Trigger Chains

A completely new system for building reusable, multi-step post-processing pipelines:

• Visual flow builder with drag-and-drop nodes
• IF/ELSE conditional blocks for branching logic
• Parallel execution groups for concurrent processing
• Scheduling — run trigger chains on a schedule
• Run history with detailed execution tracking
• Webhook triggers — trigger chains from external services
• Checkpoint and resume support

Trigger Chain Features

Feature	Description
Visual Builder	Drag-and-drop interface for building trigger chains
IF/ELSE Blocks	Conditional branching based on expressions
Parallel Execution	Run multiple triggers concurrently
Scheduling	Run trigger chains on a schedule
Run History	Detailed execution tracking with retry support
Webhook Triggers	Trigger chains from external services via HTTP
Checkpoint & Resume	Resume failed chains from where they left off

Engine MCP Server

The rtila-engine binary now doubles as a full Model Context Protocol (MCP) server when started with --mcp. This exposes RTILA’s browser automation, page analysis, project execution, and data access to any MCP-compatible client.

Key capabilities:

• 17 tools across three categories: Browser Tools (async), Run Control, and Data Tools (sync)
• Async execution model — browser tools return a run_id immediately; clients poll for completion
• Direct database access — data tools read projects, datasets, scraped rows, and logs directly from PocketBase
• Read-only resources and built-in prompts for clients that support them
• Cross-platform — works on Windows, macOS, and Linux
• Client support — tested with Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Continue.dev, and Cline

Configuration:

Flag	Environment Variable	Purpose
--mcp	—	Enables MCP server mode
--rest-api <url>	RTILA_API_URL	Database API URL
--rest-api-token <token>	RTILA_TOKEN	Auth token (sent raw, no Bearer prefix)
--project-id <id>	RTILA_PROJECT_ID	Default project
--session-id <id>	RTILA_SESSION_ID	Default session for runs

See MCP Server for complete documentation.

Removed Features

Feature	Reason
RAG / Qdrant	Replaced by full-context injection. See Breaking Changes.
Finetuned Models	Replaced by any GGUF model. See Breaking Changes.
Custom Provider	Merged into Local LLM → Custom API Endpoint mode.
RAG Top K setting	No longer needed. Full context is always injected.
RAG Similarity Threshold	No longer needed. Full context is always injected.
RAG Server Status indicator	No longer needed. RAG is removed.
RTILA Lite 1.5 Assistant models	Replaced by any GGUF model from Hugging Face.

Migration Guide

Old Setting	New Location	Action Required
AI Provider: “Local LLM” (finetuned)	AI Provider: “Local LLM” (GGUF)	Select a GGUF model from Hugging Face
AI Provider: “Custom Provider”	AI Provider: “Local LLM” → Custom API Endpoint	Enter your API base URL
AI Provider: “OpenRouter”	AI Provider: “OpenRouter”	No change needed
RAG Top K / Threshold	Removed	No action needed (RAG is gone)
Context Size: 4K–128K	Context Size: 4K–1M	Default changed to 32K
RTILA Lite 1.5 Assistant	Any GGUF model	Select a model like Qwen 2.5 7B or Llama 3.1 8B
RTILA Lite 1.5 Assistant Q4	Any GGUF Q4 model	Select a Q4 quantized model
RTILA Lite 1.5 Assistant Q8	Any GGUF Q8 model	Select a Q8 quantized model

Bug Fixes

• Fixed KV cache memory leak during long conversations.
• Fixed model download progress reporting for large files (>10 GB).
• Fixed llama-server port collision when multiple instances were running.
• Fixed context truncation not preserving critical command reference sections.
• Fixed error toast not displaying hardware diagnostic messages.
• Fixed browser profile data not being properly isolated between profiles.
• Fixed trigger chain checkpoint data not being saved correctly.
• Fixed webhook signature verification for payloads with special characters.

Performance Improvements

• 50–70% faster first-response latency for local LLM (KV cache eliminates re-processing of system prompt).
• 40% reduction in RAM usage by removing Qdrant + embedding model.
• Zero RAG retrieval latency (full context is always available).
• Faster build times (no Qdrant binary download, no Python venv for ingestion).
• Reduced disk usage (no Qdrant storage, no embedding model files).
• Improved model loading (KV cache optimization reduces model load time).

Version 8.3.151 (Previous)

New Features

• Smart Network API Interception — Automatically intercept API responses to auto-build datasets.
• Full Telegram Bot Integration — Control RTILA from your phone via Telegram.
• Isolated Browser Profiles — Isolated browser environments with unique fingerprints.
• Node-Based Integration Canvas — Visual flow editor for triggers.
• Upgraded Visual Recorder — Improved recording with loop support.
• Refreshed Workspace UI — Updated interface with improved navigation.

Improvements

• Enhanced AI self-correction with detailed logging.
• Improved humanoid mouse movement simulation.
• Better proxy rotation and failure detection.
• Enhanced checkpoint and resume system.
• Improved element picker with better selector generation.

Bug Fixes

• Fixed browser launch failures on some systems.
• Fixed memory leaks during long-running automations.
• Fixed selector generation for dynamic elements.
• Fixed proxy authentication handling.
• Fixed checkpoint data corruption on crash.