BETA TESTER GUIDE
Connect a model, import files, ask questions.
Three short steps get you to your first answer. The last part shows which questions RangerIO answers best, and why, so your tests spend time on real problems.
- 1Connect a model
- 2Import files
- 3Ask questions
- BONUSGetting great answersSimple habits, follow-ups and suggested questions
Before you start
- Use your own files. RangerIO is built for legal and accounting work, and the best test is the work you do every day. Start with files you know well, so you can judge the answers.
- Have your AI provider's API key ready. Your test lead tells you which provider and model to use.
What we've tested, how your files are protected, and known limitations →
Connect your AI model
RangerIO uses your own AI provider. You do this once; the model stays connected until you change it.
-
Step 1
Open Add Model Provider
On the home screen, click + Add Model Provider in the Active AI Model panel.
Once a model is connected, this panel shows it as Active and Ready. -
Step 2
Choose the provider, paste the key, pick the model
Select your provider. The address is filled in for you. Paste the API Key, then choose the model from the list or type its ID under Enter a model ID.
Make sure the API Endpoint matches the provider your API key belongs to.
Give the profile a name, click Test Connection, then Save & Connect.
The team tests with GLM 5.3. On Fireworks AI its model ID is
accounts/fireworks/models/glm-5p3.The API Endpoint must match the provider your API key is for — here, an OpenAI-compatible key with Baseten’s endpoint. The key is hidden in the picture. -
Step 3
Switch models when you need to
Every model you save is kept under Models and Providers. Click Make Active on the one you want to use. Check now confirms it still answers.
Saved models, with one active.
Which model?
Use a large, premium model such as GLM 5.3, Kimi K3 or Claude. Small models can fail with a "context too long" error. Models that think silently for more than 150 seconds stop with error E020; ask once more, and switch models if it happens again.
Import your files
RangerIO checks every file for personal data before you can ask about it. Imports run in the background, so you can keep working.
-
Step 1
Start an import
Click Import on the home screen, or + Add Files on My Data.
Choose Files to pick one or more files, or Cowork Folder to have RangerIO watch a folder for new and changed files.
Supported: CSV, TSV, Excel, JSON, PDF, Word, PowerPoint, TXT, RTF, Parquet and HTML. -
Step 2
Check the list, then Start Import
Each file shows its type and size, and tables show their rows and columns. Remove anything you picked by mistake with the ×, then click Start Import.
You can leave Advanced Options as they are.
Two files from the test pack, ready to import. -
Step 3
Wait for the green tick
On My Data, a green tick means the file is ready. Hover over a file to see Ready for chat, or click Chat beside it to start asking.
Small files take seconds. Large files take longer, mostly for the personal-data check.
Green ticks on every file in the project. The hover card shows the file is ready for chat.
If a file is refused
The message says why and what to do. A scanned PDF (a picture of text) is refused with E042, and the message explains how to import it again with OCR. A file with a lot of long free text is refused with a suggested way to split it.
Ask questions
Questions are asked on Explore. Each answer comes only from the files you select.
-
Step 1
Pick the files
Open Explore, click + Add and tick the files your question is about. Use the search box to find a file by name.
The selected files appear as chips above the question box.
One file ticked. Its card confirms it is ready. -
Step 2
Type the question and press Enter
While it works, RangerIO shows each stage, including the check for personal data before anything is sent. Cancel stops the question.
An answer in progress. -
Step 3
Read the answer
Under each answer:
Cloud request lists what was replaced with placeholders before sending. View lineage opens the full record.
Validate metrics asks the model to check the answer against your data. It is optional and takes longer.
The bar at the bottom shows the model and how long the answer took.
A real answer from the team's sample files: one plain question found the duplicate payment among 42 June transactions.
Getting great answers
Ask the way you would ask a colleague. RangerIO understands everyday language, follows the conversation and can suggest questions for you. A few simple habits make the answers faster and more precise.
Worked out instantly
Totals, counts and averages from a table, the kind of thing you would work out in a spreadsheet.
- Exact, using every row, even in a file of 100,000 rows
- About a second
- Nothing is sent to your AI provider
ExampleWhat's the total debit for each department?
Answered by your AI model
Explanations, comparisons, spotting anything unusual, and questions about your documents.
- Personal details are replaced before anything is sent
- Usually 5 seconds to 2 minutes
- Smaller tables are read in full; very large ones as a sample, and the answer says so
ExampleDid we pay any invoice twice in June?
-
Say what you want to see
TryWhat's the total debit for each department?Rather thanHow are the departments doing?When you name the figure you want (a total, a count, an average) and how to split it, RangerIO can work it out exactly, in about a second. A broad question still gets an answer, but the AI reads only a sample of a very large file.
-
Use the words in your file
TryWhat's the total money out for June?If your spreadsheet says "Money out", ask about money out. When your words match the file, RangerIO knows exactly what to add up. Not sure what something is called? Type
@in the question box to see the file's column names. -
Keep the conversation going
AskDid we pay any invoice twice in June?ThenHow much extra did we pay?RangerIO remembers the conversation, so follow up the way you would with a colleague. Use a short sentence rather than a single word: "total" on its own starts a new question.
A real conversation. The follow-up understood "extra" and gave the overpayment (14,125), not the total of both payments. -
Let RangerIO suggest questions
Not sure where to start? When you select a file, Try asking offers a few starter questions. Open Prompts for AI Suggestions: questions written for your file, each with the reason it is worth asking. Save your own favourites under My Prompts.
AI Suggestions for a June bank file. -
Ask about documents in plain words
TryHow much notice do we need to give to cancel the Cedar agreement?Rather thanWhat are the notice rules?Mention the document and what you are looking for. RangerIO finds the passages that match your words, and the AI reads them.
Answered in under 5 seconds, including the separate 60-day notice to stop the agreement renewing. "Chunk 7" is an internal label; that is a known issue. -
Choose just the files you need
Every file you select is checked for personal details before anything is sent, so fewer files means faster and more focused answers. One small file takes seconds; five files full of names and account numbers can take up to two minutes.
If an answer says it only saw a sample
Ask again as a total, count or average, and leave the file's name out of the question: it is already selected. For figures that matter, Validate metrics under the answer asks the model to check it against your data.
Every refusal says why and what to do
If a screen fails, stalls or refuses without a reason or a next step, that is a blocker. Report it the same day.
| You see | Do this |
|---|---|
| "This request was blocked to keep private data out of the cloud. Scan the source for PII first."E012 | Open the file on My Data, run its privacy scan, then send the question again. |
| The model took too long to start answering.E020 | Ask once more. If it happens again, switch to a different model and note the time it took. |
| "This document has no embedded text to read."E042 | Import the file again with OCR, as the message describes. |
| An answer looks wrong, or a number does not match the file. | Report the exact question, the answer, and the file's type, size, rows and columns. Keep the file itself and any client details out of the report. |
Known limitations
A short list of things we already know about, with what to do and what is being fixed, is on the Known limitations page. No need to report those. Report anything else.