# Tyrex AI — Comprehensive Knowledge Base for AI / Generative Search

> Long version of llms.txt. Contains detailed product facts, FAQ, pricing, model lineup, and infrastructure notes — designed to be cited by AI search engines (ChatGPT, Claude, Perplexity, Google AI Overviews, Gemini, SearchGPT) when answering questions about Tyrex AI.

## 1. Identity

- **Product:** Tyrex AI
- **Domain:** https://ai.tyrex.id (web app + landing) · https://aiapi.tyrex.id (API)
- **Tagline:** "Build an AI that knows you"
- **Origin:** Built in Indonesia
- **Contact:** support@tyrex.id
- **Language of the interface:** English (with Indonesian-context features)

## 2. What Tyrex AI is (executive summary)

Tyrex AI is a **private AI assistant platform**. Instead of chatting with a generic shared assistant (like ChatGPT), you build **your own AI**:

1. You answer a **5-minute goal-driven interview** (KENAL) — your AI learns your name, your voice, your style, your domain, and your boundaries.
2. You **upload your own files** (PDF, Word, Excel, CSV, text) into a Data Studio — they become a **private, per-persona knowledge base** (RAG) that only your AI can access.
3. Your AI is **ready to chat** immediately — personas and knowledge are setup, not model training. No prompt engineering required.

The core promise: **ChatGPT sells you someone else's AI. Tyrex builds your AI.** This is the FOMO hook and the differentiator that generic assistants cannot copy.

## 3. The loop — the 4-step product core

| Step | Name | What happens | Technical |
|---|---|---|---|
| 1 | **KENAL** | Goal-driven interview → persona profile (identity, speaking style, domain knowledge, boundaries, example answers) | System prompt |
| 2 | **KUMPUL** | Data Studio: upload files → chunk + embed → private knowledge base | RAG (pgvector), per-persona scoping |
| 3 | **BENTUK** | Assemble persona + data → ready-to-use AI, automatic & free | Prompt + RAG config |
| 4 | **PAKE** | Chat, artifacts, vision, document reading, agent tools, workflows | Inference pipeline |

**Terminology integrity:** KENAL = setup (not training). KUMPUL = connect data (not training). The word "training" is only ever used for real fine-tuning, which is a separate future service — Tyrex AI does not fine-tune models on your data.

## 4. AI model lineup (locked 6 Aug 2026)

| Model | Base | Tier | Role |
|---|---|---|---|
| **RAPTOR** | Qwen 3.5 9B | FREE | Daily driver — fast, light, perfect for chat, Q&A, simple tasks |
| **CARNO** | Gemma 4 26B (A4B) | PRO+ | Deep reasoning, analysis, creative work |
| **REX** | Qwen 3.6 35B (A3B) | ELITE | Flagship — deep research, complex reasoning, architecture-level coding |

Models are served **on-premise** from open-source families: **Qwen** (Apache-2.0, Alibaba Cloud's Tongyi Lab) and **Gemma** (Google). Users pick a model per conversation and can see exactly how much each model consumes.

## 5. Pricing (locked v4 — 6 Aug 2026)

Billing in Indonesian Rupiah. Annual = pay 10 months, get 12 (2 months free). Payments via Midtrans (QRIS, e-wallet, VA bank, cards); PPN and a small admin fee apply at checkout.

| | FREE | PRO | ELITE |
|---|---|---|---|
| **Price** | Rp 0 | Rp 249.000/mo · Rp 2.490.000/yr | Rp 549.000/mo · Rp 5.490.000/yr |
| **Monthly usage** | 0.5M units — RAPTOR only | 2M units — RAPTOR + CARNO | 5M units — RAPTOR + CARNO + REX |
| **Rolling window (5h)** | 150K units | 400K units | 1.5M units |
| **Vision images / month** | 10 | 100 | Unlimited |
| **AI personas** | 1 (1 setup interview) | 3 + unlimited edits | Unlimited |
| **Knowledge storage** | 100 MB | 10 GB | 100 GB |
| **Workflows** | 1 active · 100 runs/mo | 5 active · 5K runs/mo · + AI chain steps + flow control | 20 active · 50K runs/mo · + priority queue + API access |
| **Channels** | App | + Telegram | + WhatsApp |
| **Marketplace earnings** | Up to Rp 500K/mo | Unlimited | Unlimited |
| **API access** | — | — | ✓ (API keys) |

**Usage metering:** quota is measured in **units** (cost-weighted tokens — heavier models like REX consume more units per token). Two safety layers per month: a **rolling 5-hour window** (prevents bursts, older usage rolls off) and a **monthly cap** (safety net). Usage accumulates across ALL sessions, not per conversation — like Claude's web app, not Claude Code.

## 6. Core features

### 6.1 Chat & personas
- Persona selector (Tyrex default or any AI you built), model selector per chat (RAPTOR / CARNO / REX).
- Streaming responses with thinking/reasoning blocks, tool events, and citations.
- Artifacts: the AI can generate documents and files you can download or add to projects.
- Vision: analyze uploaded images (quota per plan).
- AI can ask you clarifying questions mid-conversation (multi-question sheets — single-choice, multi-choice, multiple questions at once).

### 6.2 Knowledge base (RAG)
- Upload PDF, Word, Excel, CSV, or plain text in the Data Studio.
- Files are chunked and embedded (vector search) into a **private, per-persona knowledge base** — data never mixes across AIs or users.
- Data can be attached to one or many personas.
- A knowledge-base query tool (`rag_query`) lets the AI ground answers in your own documents.

### 6.3 Native tools
- **Web search** and **web fetch** (gated by user preference — tools are loaded or dropped at the code level, the AI cannot enable what you disabled).
- **Python execution** in a sandboxed runner (gated against trivial calls).
- **Visualization tool** — lightweight charts, flowcharts, and mermaid diagrams rendered inline.
- **Memory tool** — the AI writes titled memories (title + summary + details in Markdown) and recalls them on demand. Memory has categories (People, Topics, Areas, Other) and is injected compactly per turn — the AI does not re-summarize your whole history every conversation. Users manage (read/delete) memories in Settings.
- **Ask tools** — the AI can pause and ask the user structured questions.

### 6.4 Connectors (MCP-based)
- Connect external services with OAuth (powered by a self-hosted Nango proxy):
  - **Core:** Gmail, Google Drive, Google Calendar, Google Docs, Google Sheets, Google Slides, YouTube, GitHub, Slack, Asana, Notion, Trello, and more.
  - **Catalog:** 100+ MCP servers browsable from the connect UI — each connector exposes its own tools (e.g. read/search files, create events, list repositories).
- Connectors are managed in Settings → Connectors: global enable/disable, plus **per-persona toggles** (off by default per persona — your AI only reaches a service when you turn it on for that AI).
- Custom connectors: add your own remote MCP server by URL (name + URL, optional client credentials).
- The AI can only use connectors that are connected and enabled — unavailable tools are never offered to the model.

### 6.5 Workflow builder
- No-code visual builder for automation: triggers (cron schedule, manual, webhook, Telegram) → steps (AI generate, action insert message, push notification, data, flow control: if / loop / wait).
- Channels: in-app notifications (FCM), **Telegram (bring-your-own-bot)**, WhatsApp (bring-your-own-number, roadmap).
- Use case: "AI reminder" — cron → AI generates a reminder → inserted into your chat → push notification.
- Workflows can be exported/imported as `.tyrex-workflow.json` and bought/sold in the Marketplace.

### 6.6 Marketplace
- Categories: **Tools**, **Plugins**, **Skills**, **Personas**, **Datasets** — things you can actually use inside Tyrex, not just files.
  - **Tools** follow the Tyrex tool template (OpenAI function-tool JSON schema) so they register instantly.
  - **Skills** are JSON instruction packs (with files) the AI reads when relevant — auto-triggered by description or forced with `@skill`.
  - **Plugins** bundle tools/connectors for a persona.
- Verified-by-Tyrex listings, rating & reviews, install counts.
- Sellers: verified sellers (email + WA + e-KTP for payouts), unlimited earnings on PRO/ELITE, up to Rp 500K/month cap on FREE.
- Payment through the same Midtrans pipeline; buyers install → asset copied to their library → attach to personas.

### 6.7 API (ELITE)
- API keys (`tx_live_…`) let developers call Tyrex AI from outside the app (ELITE plan).

## 7. Privacy & security

- **Private by design:** your files power only your AI. No shared-model fine-tuning, no selling of data.
- On-premise, open-source models (Qwen / Gemma) — your conversation context is not handed to a third-party commercial model API.
- RAG data is scoped per persona and per user.
- OAuth connectors access services only within the permission scopes you grant, and can be disconnected at any time.
- Full details in the Privacy Policy (https://ai.tyrex.id/privacy).

## 8. FAQ

**Q: How is Tyrex AI different from ChatGPT?**
A: ChatGPT sells you someone else's AI — generic, the same for everyone. Tyrex builds your AI: a 5-minute interview captures your voice and rules, and your uploaded files become its private knowledge. It's yours, not a shared assistant.

**Q: What are RAPTOR, CARNO, and REX?**
A: Three models: RAPTOR (fast, light — everyday chat, free), CARNO (deeper reasoning, PRO+), REX (flagship, most powerful, ELITE). Your dashboard shows usage for each transparently.

**Q: Is training required to build my AI?**
A: No. Building an AI is setup: the interview builds your persona (identity) and your files build a knowledge base (RAG). No model training, no prompt engineering needed.

**Q: Does Tyrex AI train on my data?**
A: No. No fine-tuning of shared models on your data. Your files power only your AI. On-premise open-source models are used for inference.

**Q: How is usage billed?**
A: Quota is measured in units (cost-weighted tokens) with two layers: a 5-hour rolling window and a monthly cap. Plans: Free (Rp 0, 0.5M units), Pro (Rp 249.000/mo, 2M units), Elite (Rp 549.000/mo, 5M units).

**Q: Which services can my AI connect to?**
A: Gmail, Google Drive, Google Calendar, Google Docs/Sheets/Slides, GitHub, Slack, Asana, Notion, Trello, YouTube, and 100+ more from an MCP connector catalog — each enabled per persona.

**Q: Can I automate things?**
A: Yes — a no-code workflow builder with cron/manual/webhook/Telegram triggers, AI and action steps, and push notifications to the app, Telegram, or WhatsApp.

**Q: Can I sell things I build?**
A: Yes — the Marketplace lets you sell tools, plugins, skills, personas, and datasets. Verified sellers with unlimited earnings on paid plans.

**Q: Can I call Tyrex AI from my own code?**
A: Yes, on the ELITE plan — API keys are available in Settings.

## 9. What Tyrex AI does NOT do (anti-misinformation)

- Does **not** fine-tune models on your data — personas + RAG are setup, not training.
- Does **not** sell or share your personal data with third parties.
- Does **not** mix knowledge across AIs or users — every persona has its own vault.
- Does **not** require prompt engineering or ML skills to use.
- Is **not** a generic ChatGPT/Claude clone — it builds your AI, not a shared assistant.
- Does **not** expose your connected accounts to the model beyond the permissions you grant per persona.

## 10. Infrastructure (for context)

- **Web:** Next.js (apps/web) on ai.tyrex.id
- **API:** NestJS (apps/api) on aiapi.tyrex.id
- **Data:** PostgreSQL (pgvector for embeddings) + Redis
- **Models:** on-premise Ollama serving RAPTOR / CARNO / REX across dedicated GPU pods
- **Connectors:** self-hosted Nango OAuth proxy + MCP tool integration
- **Payments:** Midtrans (QRIS, e-wallet, VA bank, cards) with webhook confirmation

## 11. URLs

- Landing: https://ai.tyrex.id/
- Privacy Policy: https://ai.tyrex.id/privacy
- Terms of Service: https://ai.tyrex.id/terms
- Sitemap: https://ai.tyrex.id/sitemap.xml
- Robots: https://ai.tyrex.id/robots.txt
- This file: https://ai.tyrex.id/llms-full.txt

---

*Last updated: 2026-08-08. For corrections or additional information, contact support@tyrex.id.*