AITransX model workbench guide
Test chat, image understanding, document questions, web search, image generation and speech in your browser. Actual model and tool calls are billed by usage.
Get started
- Sign in, open Model use and select an active API key. Browser calls use that key's balance, permissions and usage limits too.
- Choose Chat, Image or Speech, then select a suitable model. Available models, vision capabilities and provider routes depend on the catalog and your account configuration. Browse current models
- Enter a question and add images or files if needed. Enable web search when you need external information, and check attachments and settings before sending.
- After sending, review the answer, sources and charges. Start with a small task to check the result and billing unit before processing larger material. Check current prices
Focus mode is for continued reading and conversation. After collapsing its top toolbar, use the restore button beside the composer controls to bring back model switching and other actions.
Choose a workflow for your task
Chat, image understanding, search, image generation and speech do not have to use the same model. Check each task's capabilities and charges first.
| Your task | Where to start | What to check before sending |
|---|---|---|
| Questions, writing and organizing material | Select an answer model in Chat | The answer model's input/output rates and the context included in the request. |
| Image, screenshot and chart questions | Attach images in Chat and check vision settings | Use a vision-capable model. The answer model may read directly, or a separate vision model may read first. |
| TXT, CSV, Word and PDF questions | Add files in Chat | Review the extracted material. Scanned PDF pages need image reading; not every format is a native model attachment. |
| Current information or public web pages | Enable web search in Chat | Planning, search, page reading and the answer may have separate charges. Review the cited sources afterwards. |
| Generate images or use reference pictures | Switch to Image | Supported reference images and parameters, the billing unit and downloading the result. |
| Turn text into speech | Switch to Speech | The model's supported voices, formats, languages and billing unit. |
Images and document questions
Choose a vision model and explain your question
After adding images, check vision settings and search or filter the available vision models. The workbench shows the current default; not every chat model supports images natively. Your answer model may read images directly, or a separate model may pass a reading summary to it. The page identifies the actual workflow.
Follow-up questions use the previous answer or image-reading summary; they do not guarantee another read of the original image. Attach the image again and specify the area when checking small text, colors or other details.
Ordinary documents are read in your browser first
The browser extracts text and presents an attachment before selected material is added to the model input when you send. The answer model does not need native Word or PDF upload support for this workflow. Extracted text still uses model context and is billed as actual input usage.
Currently supported document formats include: TXT / MD / LOG / JSON / CSV / TSV / DOCX / PDF
For a scanned PDF without extractable text, choose the pages to read as images. Convert older Word (DOC), spreadsheets and other unsupported formats first. The current page validation determines file size, page and text limits.
Local reading does not mean local model inference. After you send, AITransX forwards the selected text or images to the chosen upstream. A local conversation is not a promise to retain the original file permanently. Read the privacy policy
Web search and reference sources
Enable search to check current information, public facts or web pages. Editing text or extracting information from your own attachments often does not need an extra search.
- Enable web search and ask a complete question. You usually do not need to prepare keywords: the system decides whether the question needs search planning.
- Use search settings to restrict sources to particular sites. You can search, filter and customize the planning model; check the current settings for its default selection.
- When needed, the workflow reads a small number of pages before the answer model summarizes the material. Citation numbers and source lists let you check the original web pages.
You do not need web search just to read an attached document. Cost details distinguish the planning, search, page-reading and answer calls that actually occurred.
Public pages may still be incomplete or inaccessible because of sign-in requirements, blocking, expired links or extraction failures. Citations help verification; they do not guarantee accuracy or complete coverage.
Generate images and speech
Generate pictures in Image
Choose an image model and describe the subject, style and composition. Add reference images when the model supports them. Image generation is a separate task; adding a picture in Chat is mainly for understanding or analysis, not a request to generate a new image.
Synthesize text in Speech
Select a speech model, choose a supported voice and parameters, enter text and generate audio. Voices, formats and other options vary by model; another model's parameters are not a universal preset.
Download images and audio you want to keep. Local image history helps revisit saved results, but browser storage, temporary links and manual cleanup can affect later access.
Local history, backups and sharing
Chat history stays in the current browser. Signing in on another device does not automatically bring it across. Account preferences and conversation content are different data; preference sync is not a cloud conversation backup.
- Select the records to include in later chat from local conversations. Keeping a large history does not mean every saved message is sent with every request.
- Export important conversations and separately download images or audio you need. Check your backups before changing browsers or clearing website data.
- Persistent storage depends on the browser. It can reduce automatic eviction, but cannot prevent manual clearing of website data or replace backups.
- Storage information separates records from image and attachment usage. Cleanup keeps items still referenced by conversations or drafts, and items being saved. Read the cleanup result when checking remaining usage.
Review what you make public
Share only the messages you choose. Original attachment text is omitted from the public share, but an AI answer may quote it; review selected answers and images first. Anyone with the link can access the share page until deletion or expiry under its stated rules.
Charges and recovery
Selecting, previewing or reading attachments locally does not itself start model billing. Sending starts the model and tool workflow.
A task can include an answer, image reading, search planning, search or web-page reading. Actual calls are billed separately. Image generation and speech use their models' published billing units. Check current prices
Stopping or a later step failing does not automatically make earlier completed calls free. Check charges and request records. If a cost is unconfirmed, keep the request ID when contacting support.
- Keep your question, attachments and request ID, then read the error details. Check the key, balance, model access and selected parameters.
- For upstream disruption, follow the page's advice to wait or choose another available model, and check service status. Check service status
- A timeout or leaving the page does not prove the upstream did not execute. Check recovered results and logs before retrying, especially for billable operations such as image generation.
Frequently asked questions
Can every chat model understand images and read files?
Not every model natively supports images. Ordinary documents are converted to text locally. Images and scanned pages need an available vision model, subject to the model and route configuration.
Does local file reading mean my content stays on my computer?
Reading before sending happens locally. After you send, selected text or images enter the model workflow. Only add material you have permission to process, and review it before sending.
Does leaving the page cancel a request I already sent?
Do not treat navigation as cancellation. A sent task may continue. Use Stop, then check recovery status, results or logs when you return to avoid sending the same request twice.
How does the browser workbench differ from the API?
The workbench provides model selection, local file reading and result controls. The API lets your application send requests directly; browser file reading and multi-step tools are not a single universal API parameter.
Verify a small task before integrating
Once the model's capabilities, sources and charges work for you, reproduce the request using the API documentation. Check routes, limits and recovery before production.