Find out whether it works on your documents
RAG Playground: Ask, and see where the answer came from
Everyone wants an assistant that knows their own handbook, contracts or product notes. This is where you find out whether that would actually work — on your documents, with your questions, before anybody builds anything.
Reading your files
Done — 412 passages ready to be searched.
You asked
“How much notice for annual leave?”
Four weeks for anything longer than five days. Underneath the answer sit the two passages it was built from, so you can check it in seconds.
- Passages
- 412
- Sources used
- 2
- Platform
- In your browser, at rag.trustworthlabs.com
- Audience
- Teams deciding whether an assistant on their own documents is worth building
- Scope
- Upload, ask, compare — with the source shown under every answer
- Positioning
- A bench for trying the idea out honestly, not another chatbot to sign up to
What it does
Answers you can check
The usual demo is somebody else’s documents answering somebody else’s questions, which tells you nothing. Here it is your material from the first minute, and every answer arrives with its receipt.
Your own documents, straight in
Drop in the handbook, the contracts, the product notes — PDFs, Word files, spreadsheet exports, web pages, plain text. They are read, split into passages and ready to be asked about while you are still on the page.
Every answer shows its working
Under each answer are the passages it was built from, strongest first. When an answer is wrong you can see immediately which document misled it, which is usually the real problem.
Change one thing and ask again
Put the same question to a different AI, or index the documents a different way, and compare the two answers side by side. On your material, not a demonstration set.
It admits when it doesn’t know
You write, in a sentence or two, what the assistant is for and what it should do when your documents don’t cover the question. Saying “I don’t know” is the default, not an afterthought.
Your experiments are kept
Each setup and the conversation that went with it is saved. Come back a week later, open the one that answered best, and carry on from there.
You can see what it uses
A running count of files, passages and usage sits at the top of the screen, so the cost of the thing you are considering stops being a guess.
An afternoon with it
From a folder of documents to a straight answer about whether this works
The order below is the one the screen is built in, and it is deliberate: nothing is asked of the AI until your own documents are in and you have said what it is for.
- Sign in and drop in the documents you actually care about.
- Say how they should be read and where the index should live — or take the defaults and move on.
- Pick which AI answers: OpenAI, Google, Anthropic or DeepSeek.
- Write, in a sentence or two, what it is for and what it should say when it doesn’t know.
- Ask the questions your team would really ask, and read the passages under each answer.
- Change one piece, ask the same question again, and keep the setup that answered best.
Your documents
Where what you upload actually goes
Your documents are used for one thing: answering your own questions in your own sessions. They are not shown to other accounts and nothing is trained on them. You can remove a file from a session whenever you like, and the passages that came out of it go with it.
One thing worth saying plainly, because it is true of every tool of this kind: to answer a question, the relevant passages are sent to the AI provider you picked — OpenAI, Google, Anthropic or DeepSeek. Choosing the provider is choosing who sees that text. It is a playground, so use documents you would be comfortable handing to them.
If you want to run the same thing on your own infrastructure instead, with nothing leaving your network but the model call, that is a conversation worth having — ask us.
Under the bonnetHow it is built. Only interesting if you work in software — nothing here changes what the product does for you.
Retrieval-augmented generation, assembled a stage at a time
Kept deliberately short — none of this changes what the playground does for you.
- One question, five stages
- Files are split and indexed, the relevant pieces are found, and only those are handed to the AI with your instructions. The screen is laid out in that order, so what you build is what runs.
- Four AI providers, one interface
- OpenAI, Google, Anthropic and DeepSeek all sit behind the same question, and switching between them takes one click.
- Your choice of where the index lives
- Several search databases are wired up, so a setup you settle on here is one you can rebuild on your own infrastructure.
Try it on your own documents before you commit to anything
The playground is open. If what you find there is worth building properly — on your infrastructure, wired into the systems you already run — that is the part we do.