What companies use it for
Questions into contracts and policies
Notice periods, penalties, payment terms — with a page reference, so it can be quoted in a meeting.
Onboarding a new person
A new hire works through procedures and policies; when something isn't clear, they ask the assistant and get the answer from the current manual.
Reports from company data
How much was sold, or when the last order came in — without writing database queries.
Support on the website
The assistant answers customers from your documents; Shield guards what comes into it while it does.
Ask your documents. Get an answer with a citation.
Studio answers from your documents and always shows the source. The demo knows three sample documents of a fictional company — ask them anything.
Production Studio: hybrid search (lexical + vector) over your knowledge base, answer via your chosen LLM (OpenAI / Anthropic / Google / a local model) with document-and-page citation; PII redaction at ingest. This demo uses scripted answers over a fixed set of three documents — no LLM call.
The demo runs entirely in your browser and sends nothing anywhere. Even so, please don't paste real sensitive data.
An answer is built in four steps
Documents are prepared in advance, and for each question only what relates to it is pulled out.
- 01Ingestion
Documents are split into chunks and indexed. Every uploaded document is first checked for hidden instructions aimed at the assistant.
Technical detailWhen an uploaded document is processed, an antivirus scan and hidden-instruction detection (knowledge-base poisoning) run. A document flagged as malicious or carrying a hidden instruction never makes it into the index, and so it never shows up in answers. If the scanner isn't available, the document is processed anyway.
- 02Retrieval
Relevant passages are found for the question; questions over data run read-only.
Technical detailHybrid search combines lexical and vector matching; questions over company data run as a read-only SQL query.
- 03Answer
The model you choose receives only the found passages and the question — not the whole database.
Technical detailYou choose the model; only the found passages and the question are sent to it.
- 04Citation and record
The answer carries a citation of its source; the conversation is saved, and sensitive-data findings stay in the record.
Technical detailA citation is shown only if the passage passes a relevance check: a greeting or small talk doesn't trigger retrieval at all, and loosely related passages are dropped before the answer.
What Studio is made of
Studio isn't just a chat: four parts, each with a different job.
Knowledge base with citationsKnowledge base with citations
Documents, spreadsheets and web pages together. The assistant answers from them, cites the source, and says so when an answer isn't backed by the material.
Sources: PDF, Word, Excel, CSV, PowerPoint and web pages. On ingestion a document is split into parts and indexed; before that it passes an antivirus scan and hidden-instruction detection. Search is hybrid, lexical and vector, and a citation is shown only if the passage passes a relevance check. In the cloud deployment, before delivery the answer is verified against the retrieved passages by the same grounding verifier Shield offers to third-party applications.
Chat over company documentsChat over company documents
Nobody needs to know which file to look in. You ask in plain language and the answer comes with a citation.
When the model runs in the cloud, sensitive data from the question and the retrieved passages leaves only as placeholders and the employee sees the real values in the answer; how the placeholder map is kept across the whole context is described below under “Boundaries enforced in code”. With a local model the data never leaves your infrastructure and this step is skipped by default.
Analysis moduleAnalysis module
Questions to company data in plain language — how much, when, which. The data is only read; the result is a table or summary, not a guess from the model.
A connected database is only read: a query must start with SELECT or WITH, and write or schema-changing statements are forbidden; for PostgreSQL the connection runs in a read-only transaction with a statement timeout. The result is a table or summary drawn from the data, not a guess from the model.
Where it runs and which modelWhere it runs and which model
In the EU cloud, on your own server or fully local — as you choose.
The model can come from a public provider such as OpenAI, Anthropic or Google, or run on your own server — in combination, too. The choice is made at every step, so a step involving personal data can stay with you even when the cloud is allowed. With your own key to the model, we don't bill the tokens.
An assistant without a source is just a persuasive guess. The citation is what turns an answer into something you can take to management or a client.
The assistant improves from how people reactAuto-learning: from reactions to in-context examples
A correction or a satisfied reaction isn't lost: it becomes an example for a similar question.The model's weights are not fine-tuned. Conversations produce training examples that are inserted into the system prompt.
- 01Reaction
A thumbs up or down, a correction of their own, but also a plain “thanks, exactly” or “that's not it”.
Technical detailThe unit analysed is the triplet question → answer → reaction. Explicit feedback (thumbs up, thumbs down, correction) is complemented by implicit feedback: keywords in the supported languages and a model classifying the reaction as positive, negative, correction, rephrase or neutral.
- 02Proposed example
An example is only created from an answer the person was happy with, or from their correction.
Technical detailAn example comes only from a positive signal with a high quality rating or from a correction signal, where the user's text becomes the expected output. An exchange with critical or high sensitive-data risk is skipped, and so is a duplicate question.
- 03Review
Examples from legal, finance, HR or security, and examples with suspicious content, wait for an administrator's approval. The rest are added on their own when confidence is high.
Technical detailThe legal, finance, HR, compliance, medical and security categories always go to the approval queue. Before saving, the planted-instruction detector runs — the same one the knowledge base uses; a hit doesn't delete the example, it only removes automatic approval, and the same happens if the detector fails. Other examples with sufficient classification confidence are approved automatically.
- 04Use
An approved example is added to the material the assistant gets with the question. An administrator can edit or retire it; the kill switch, which stops outgoing data, turns learning off as well.
Technical detailActive examples are inserted as few-shot in priority order: corrections, company context, questions and answers, instructions; questions and answers are picked by relevance to the current question. A daily limit caps the flow of new examples.
Learning picks up mistakes too: a praised wrong answer can be approved on its own, and the assistant repeats it until an administrator fixes it. Look through what has been added now and then.
What it looks like in practice
Sample scenarioNot references to specific clients, but typical situations.
A company deploying an assistant over its own documents
- — Everyone searches the shared drive on their own for answers to the same questions
- — New hires ask colleagues about things that are already written in the manual
- — With an assistant, nobody knows whether it made the answer up
- A question is enough, and the answer shows the source
- Materials are uploaded once and apply to everyone
- For a question outside the material, the assistant admits it has no answer
Most of the effort went into choosing the documents that count as the source of truth.
Studio can also be made available over MCP. Details are on a separate page.
Protection for outgoing data is part of every Studio — Complete package, at no extra charge.
The assistant is only as good as the material it's given. With five versions of the same policy, someone first has to decide which one is valid — the model won't do that for you.
Boundaries enforced in code
A data query must start with SELECT or WITH, and write or schema-changing statements are forbidden. For connected PostgreSQL sources, the connection additionally runs in read-only transaction mode with a statement timeout.
With a cloud model, the system prompt with the retrieved passages, the conversation history and the new message are anonymised with a shared map, so the same person gets the same token everywhere. The answer is de-pseudonymised even when streamed. If a local model is running, this step is skipped by default.
When data protection finds something in a message, the audit stores the risk level, categories, number of findings, provider and a SHA-256 fingerprint instead of the message itself.
Let's start with measurement. You'll see what's happening in your company today.
First you see numbers from your own operation: how many bots, how much spam and what leaves for AI. What we turn on is your call after that.
Technical detailThe widget and the extension need no change to your application's code; SDK integration means adding middleware for Node.js, Python or PHP. Watching runs in monitor mode: events are recorded, nothing is stopped.