The complete RafayGen platform guide: every tool and how to use it

This is the full manual for RafayGen: every surface, what each one is for, how to get good results from it, and the limits worth knowing before you rely on it. It is written to be read start to finish once and then skimmed for the section you need.

Everything below runs in a browser at rafaygen.com behind one account. There is nothing to install.

Getting in

Three ways, in increasing order of commitment. You can continue as a guest and start a conversation immediately — those conversations are deliberately ephemeral, are not attached to any identity, and are not recoverable later, which is a privacy decision rather than a limitation. You can sign up free, which gives you saved history across devices. Or you can sign up and upgrade to a paid tier for the features with a real per-request cost.

The guest route exists because "create an account to see if this is useful" is a bad trade to offer someone who has not seen it yet.

The chat workspace and the six agents

The chat workspace is the main surface. It renders markdown, code with syntax highlighting, tables, mathematical notation, diagrams, generated images and file downloads inline, and it streams answers as they are written.

The agent picker is the most important control on the page and the one most people ignore. The six agents run on different providers and different models, so the choice materially changes the answer:

• Fast — lowest latency. For quick factual questions, small rewrites, and anything where waiting is worse than a slightly shallower answer.
• Everyday — the balanced default. General writing, summaries, ordinary day-to-day work.
• Deep — long-form generation. Essays, reports, extended explanations where length and coherence matter more than speed.
• Complex — technical and structured work. Code, data, anything with a rigid shape that has to come out right.
• Thinking — step-by-step reasoning. Maths, logic, debugging, and any question where you want to see the working rather than just the conclusion.
• Frontier — agentic, multi-step work that uses tools rather than answering in one pass.

How to choose an agent in practice

A rule that works: start with Everyday. Move to Thinking when the answer is wrong in a way that suggests it did not reason, and to Fast when the answer is fine but you are waiting too long for it.

It is worth deliberately asking the same question to Fast and to Thinking once, early on, just to see the size of the difference. It is larger than people expect, and it makes the picker stop looking like a cosmetic setting.

Routing across several providers also means an outage or rate limit at one lane degrades the service rather than ending it. If one agent is unavailable, another usually is not.

The voice orb

A real-time spoken interface to the same assistant and the same account: streaming speech-to-text, a low-latency chat lane, and text-to-speech that starts speaking the first sentence while the rest is still being written. You can interrupt it mid-sentence and it stops.

It works in English and Urdu and follows you when you switch mid-conversation. Use headphones if you can — it removes the acoustic echo problem entirely and makes interruption far more reliable.

Voice is best for thinking out loud, rehearsing spoken answers, and questions asked while your hands are busy. It is worse for anything you need to keep or edit, because you cannot skim speech.

The image studio

FLUX-based generation running on hardware RafayGen operates directly. You use it by asking for an image in the chat; the request is detected and routed.

It does text-to-image, editing an uploaded picture by describing the change, image-to-image restyling that keeps the composition, upscaling, and describing an image back as text — the last being genuinely useful for writing alt text or cataloguing a folder of pictures.

One rule that saves the most frustration: never write what you do not want into the prompt. "No clutter" produces clutter, because the model is being steered by your words rather than executing them. Describe the positive state instead.

Documents, decks and spreadsheets

Ask for a presentation, report, letter or spreadsheet and you get an actual file — .pptx, .docx, .pdf or .xlsx — that downloads and opens, rather than text to paste into Word yourself.

The engine plans the structure, writes the content, renders the file, and then re-opens what it wrote to check it: that every planned section exists, that no text block overflows its frame, that the hierarchy is intact. That validation catches structural breakage. It cannot check whether the content is true, and every number and named source in a generated document should be verified before you send it anywhere.

Prompt for the document, not the topic. Audience, length, structure and tone in the first message will save you three rounds of correction.

Urdu OCR

The OCR tool converts scanned Urdu documents, photographs of pages and image-only PDFs into searchable, copyable text and a rebuilt PDF with a real text layer underneath the image.

Urdu is one of the harder scripts to recognise — Nastaliq is cursive, deeply context-dependent, and diacritically dense — so scan quality matters more here than with Latin text. Flat pages, even lighting and 300 dots per inch or better make the difference between usable output and noise.

Always proofread recognised text before relying on it. It began as a project to digitise Urdu literature and that remains what it is best at.

Languages

English, Urdu script and Roman Urdu are all first-class inputs, including mixed within a single sentence. There is no language toggle — the language is inferred from what you typed, and answers come back in the language you asked in unless you say otherwise.

Urdu output is rendered in proper Nastaliq rather than a fallback Naskh font, and the interface handles right-to-left layout when you are working in Urdu. If you want the answer in a different language from the question, say so explicitly; input language does not imply output language.

Accounts, tiers and payment

The free tier is permanent and requires no card: multi-agent chat in all three languages, with saved history if you sign up.

Go, Plus and Extreme add the voice orb, image and video generation, document generation and higher limits. They are payable from Pakistan by NayaPay or UBL bank transfer, verified by hand — usually within a few hours during working hours. Nothing auto-renews, because there is no stored card to charge; when a month ends the account returns to free and the history stays.

Limits worth knowing before you rely on it

RafayGen does not train its own models; it routes to providers that do, and inherits their failure modes. Answers can be confidently wrong. Citations, statistics and quotations are the highest-risk output and should always be verified independently.

Urdu is markedly better handled than in most global products and is still not native-speaker quality in formal or literary registers. Voice transcription degrades with background noise and heavy code-switching. Generated images are unreliable for anything requiring real-world accuracy — maps, diagrams of real mechanisms, likenesses of real people.

It is operated by one person. Support is an inbox, payment verification is a human reading email, and there is no enterprise tier. That is worth knowing in advance rather than discovering it at an inconvenient moment.