Generative AI and LLM integration

We reduce the risk in layers. The model answers from retrieved passages of your content and must cite them. Guardrails check that claims are supported by those passages. When the sources do not contain an answer, the feature says so instead of guessing. The evaluation set measures how often answers are unsupported, and we do not release until that rate meets the threshold you agree.

For answering from your documents, retrieval is usually the better start: it keeps answers current, shows sources and respects permissions. Fine-tuning helps when you need a consistent format or style, a specialised vocabulary, or a smaller model that runs cheaply. Many systems use both. We test the options against your evaluation set and recommend the simplest one that meets the bar.

Not if we can help it, and we design to prevent it. We use enterprise model services whose terms exclude training on your data, or open models hosted in your own environment. Prompts and documents are sent only to the services you approve, logs are kept under your retention rules, and personal data is masked where the task allows.

We choose per task, based on answer quality on your evaluation set, cost per request, speed, language support and where the model is allowed to run. Options include OpenAI GPT models, Anthropic Claude, Google Gemini, Meta Llama and Mistral, through services such as Azure OpenAI, Amazon Bedrock or Vertex AI. The model sits behind one interface, so it can be replaced later.

Build cost depends on the number and state of your sources, the integrations and permissions involved, and how strict the accuracy bar is. Running cost depends on usage, the model chosen, how much text is sent with each request and caching. We estimate both, and we show how a smaller model or better retrieval can bring the running cost down.

Users can try a prototype on your content early, before the full build. What takes longer is connecting sources with the right permissions, security and privacy reviews, and reaching the agreed quality on the evaluation set. We plan those steps from the start and give you a timeline once the sources and the job are clear.

Intellectual property in the work we build for you is yours from the first commit — no assignment step at the end. That includes the code, prompts, evaluation sets, retrieval indexes and any fine-tuned model weights produced for you. Any XSIS tools that existed before stay ours and are licensed to you. Commercial models remain the vendor’s and are used under their terms, which we list for you. Your documents and data remain yours throughout.

Only if you agree. We tell you which tools we would use, how they handle code and data, and whether anything leaves your environment. If you prefer, we work without them. Code written with AI assistance goes through the same review and testing as any other code, and the engineer who submits it is responsible for it.