A chatbot that answers over your documents — and cites the source.

Answers anchored in your own data, with the source in plain view. Retrieval-augmented generation over the documentation, databases and history you already hold.

An assistant that answers questions about your contracts, procedures, incident history or product catalogue. Every answer carries the document it came from, so whoever reads it can verify instead of trusting.

The problem with answers that have no origin

A general-purpose language model answers everything with the same confidence, whether it knows or not. In casual conversation that is an annoyance; where someone will act on the answer, it is an operational risk.

The fix is not a bigger model. It is restricting what the model may use as a source and forcing it to show where what it says came from.

How we build it

• Ingestion. Documents, tables, e-mails or system records are split into searchable units, each keeping its origin.
• Retrieval. The question finds the relevant passages; only those reach the model. What is not in the corpus does not enter the answer.
• Mandatory citation. The answer names the documents used. No source, no answer.
• Permissions. Search respects who may see what. Two people with different profiles get different answers to the same question — and that is exactly what has to happen.
• Confidence threshold. When retrieval finds insufficient grounding, the system says it does not know instead of inventing.

What you get

• The assistant integrated where the team already works — intranet, internal application, Teams or API.
• A usage dashboard: most frequent questions, rate of answers without sufficient grounding, and cost per period.
• A process for keeping the corpus current, so the knowledge base does not age in silence.
• A full log of questions and answers, for audit and for improving the system with real data.

On data privacy

We define upfront what leaves your infrastructure and what does not. Some cases require a model hosted in-house; in others a commercial API under a no-retention agreement is enough. The decision is yours and informed; the architecture is designed so the model stays replaceable without rewriting the system.

Frequently asked questions

Will our documents be used to train the model?

No. The corpus is consulted at question time, not used in training. Where a commercial API is used, we contract the no-retention option.

How many documents are needed?

It works from a few dozen. What matters more than quantity is that they are current and that the valid version is identifiable.

What if the assistant answers badly?

Every answer is traceable to its source, so the error is diagnosable: either the source was wrong, or retrieval brought the wrong passage. Both are fixable. That is the difference from an answer with no origin.

← All services