Private AI Knowledge Assistants & RAG Development
AI knowledge assistants using your documents, permissions and source citations. Plan retrieval, evaluate answers and integrate the assistant into your workflows.
What we can help you build
Scorchsoft builds AI knowledge assistants that help people find and use information held in business documents and connected systems. We scope the sources, user access, answer behaviour and evaluation before choosing the technical approach.
Retrieval-augmented generation (RAG) supplies relevant source material to a model when a question is asked. It can support answers with citations to policies, product documentation or operational knowledge. It does not eliminate incorrect answers, and citations still need to point to material that actually supports the response.
Here, private means access-controlled for the intended users. It does not automatically mean an on-premises model or that no third-party processor is involved. We agree hosting, providers, processing locations and data handling explicitly.
Design choices that shape the project
Retrieve authorised evidence before generating an answer
The workflow can ingest approved material, prepare it for search, retrieve relevant passages and generate an answer with source references. Source owners need a way to update or retire documents so the assistant does not keep relying on obsolete material.
Permission checks belong in the retrieval and application layers. A user should not receive restricted passages merely because they ask the right question. The answer should also communicate when the available evidence is insufficient or contradictory.
Hypothetical example: an internal support team asks which procedure applies to a particular product version. The assistant retrieves approved documentation, cites the relevant section and asks for clarification where the version is unclear. This is an illustrative use case, not a delivered customer result.
See the RAG glossary for the definition and its existing comparison with fine-tuning.
Keep knowledge access separate from authority to act
An assistant that answers a question and an agent that changes a business record need different permissions. Reading a policy does not authorise a refund, and summarising an invoice does not approve payment.
We can connect knowledge assistance to a portal or application with an explicit action and approval route. MCP integration or conventional APIs may provide those connections, depending on the workflow.
The accompanying OpsUPLOOP concept illustrates permission-aware AI over business records. A knowledge-assistant implementation needs its own source and access model; we do not assume that one product's permission rules automatically cover another system.
Scope, delivery and ongoing ownership
A scoped project can include source inventory and ingestion, search and retrieval configuration, an answer interface, citations, access controls, evaluation data, integrations and an operating runbook. Scanned documents may need document processing before their content is useful for retrieval.
Evaluate the full system using authorised, representative questions: answerable questions, missing information, conflicting sources, outdated documents and users with different access. Measure whether retrieved evidence is relevant, whether answers are supported and whether the permission boundary holds. A single accuracy headline hides these different failure modes.
Google's RAG overview describes retrieval as a way to provide external knowledge to generative models. We assess suitable existing services alongside the application, rather than assuming a new model must be trained.
Running cost includes ingestion and updates, retrieval, model usage, storage, evaluation and support. Start with a bounded source collection and business question set. An existing search or knowledge product may already meet the requirement; a custom assistant earns its place through the workflow, access or integration needs.
How we approach the work
Define the knowledge task
Agree users, questions, sources, access and acceptable behaviour.
Prepare and connect the sources
Review quality, ownership, versioning and ingestion.
Build retrieval and the interface
Implement access-aware search, source references and answer handling.
Evaluate realistic questions
Test evidence, answer support, insufficient knowledge and permissions.
Pilot and maintain
Monitor outcomes, update sources and agree ongoing ownership.
Talk through your first useful release
Bring your current process and the outcome you want. We will help assess the options and the work involved.
Questions before you commission the work
No. Retrieval can provide relevant evidence, but the model can still misinterpret it or produce unsupported claims. Evaluation, citations and a clear insufficient-information response remain important.
The implementation should restrict access to authorised users and define where data is processed and stored. Private does not automatically mean local hosting or no external provider; those requirements must be agreed.
RAG is often a useful starting point for questions about changing source documents. Fine-tuning changes aspects of model behaviour through training examples. They can be combined; the task and evaluation should guide the choice.
Potentially. Document type, scan quality, tables and extraction requirements affect preparation. Test representative material and agree which formats the system supports.
Actions can be integrated where suitable, but require their own permissions, validation and approvals. Knowledge access alone should not confer authority to change records.
Bring a bounded set of authorised source documents, example questions, user roles and the business task you want to improve. Discovery can establish whether a product, search improvement or custom assistant is the best route.
Need help building your ideas?
Tell us where you're headed and we'll come back with our thoughts, a realistic plan and a rough cost estimate. Scorchsoft is a UK-based team of app, portal and AI developers, working in-house from Birmingham's Jewellery Quarter.