02 / GRNTSOFTWARE SERVICES
Bring your business knowledge into a useful application. We design AI assistants and retrieval systems around your users, source material and existing software.
Discuss this service01 / A GOOD FIT
Design and build AI assistants, retrieval-augmented generation (RAG) systems and custom applications around your business knowledge and the people who use it.
02 / WHAT YOU RECEIVE
A technical design covering users, source collections, permissions, model access and system boundaries. We define how documents arrive, change and are removed, and what information the application is allowed to retrieve for each user.
A scoped prototype or application with a usable interface and agreed integrations. For retrieval-augmented generation (RAG), this can include document preparation, search, relevant source passages and answers linked to their supporting material.
Representative questions and expected behaviours, including missing information, conflicting sources and access restrictions. We assess retrieval and generated answers separately, alongside response time, model usage and the effort needed to review outputs.
Agreed code and configuration handover, deployment instructions and documented dependencies. Before production, we scope monitoring, content updates, error handling and ownership. Ongoing support and maintenance are agreed as part of the delivery scope.
AN ILLUSTRATIVE SCENARIO
A support team wants to find answers across approved product manuals. A focused RAG pilot would search the current manuals, retrieve relevant passages and prepare an answer with references the support agent can open and check.
If the manuals do not support an answer, the assistant should explain the gap and route the question for review. Evaluation would include outdated documents and questions outside the approved collection.
A scoping example. The final design depends on your systems and requirements.03 / HOW WE WORK
Choose users, tasks and a bounded source collection. Agree on how the application should respond when information is missing and what users need to verify.
Prepare source material, implement retrieval where useful and connect the model to the interface. Start with representative inputs before widening the collection or integration scope.
Test normal and difficult questions against agreed examples. Inspect source selection, answer support, permissions, latency and cost; record failures rather than relying on a demonstration alone.
Resolve priority gaps, connect the agreed systems and document operation. A production rollout includes agreed access controls, monitoring and responsibility for maintaining the sources.
04 / BEFORE WE BEGIN
Retrieval-augmented generation supplies a model with relevant information retrieved from an external source, such as an approved document collection. It can support questions about company knowledge and make source references available. Its usefulness depends on source quality, retrieval and evaluation; it does not guarantee correct answers.
That depends on the task. An existing model with carefully designed instructions and retrieval may be enough for a first application. We assess the task and evaluation results before recommending fine-tuning or another approach, because each adds different data and maintenance requirements.
We agree which users and systems may access each collection, then design retrieval around those permissions. The project also needs decisions about model providers, hosting, retention and document updates. These requirements are part of scoping and testing, rather than assumptions made after the interface is built.
The main factors are source preparation, permission complexity, interface requirements, integrations and the evaluation needed for the task. Model usage, hosting, indexing and support create ongoing costs. We agree the delivery scope and identify those operating costs before implementation; maintenance responsibilities are documented at handover.
YOUR NEXT USEFUL STEP
Tell us which workflow you want to improve, the systems involved and what a useful result would look like.