RAG & document intelligence
Make policies, reports and other internal documents easier to search through a question-answering assistant. Prepare the source collection, set access rules and evaluate answers against the documents they cite.
How we can help
When important information is spread across folders and document versions, staff can spend considerable time finding an answer and checking whether it is current. Retrieval-augmented generation, or RAG, connects an assistant to an agreed collection of documents so it can retrieve relevant passages and use them when preparing a response.
Ivany Concepts builds document retrieval systems and assistants with source citations, access controls and evaluation. The quality of the document collection matters throughout the work. We establish which sources are authoritative, how updates reach the system and when an answer should be referred to a person. Citations help users check a response; they do not remove the need to evaluate accuracy.
When this may be useful
- Staff repeatedly search policies, guidance and reports to answer the same internal or customer questions.
- A document collection contains overlapping versions and users need a clearer way to find the relevant source.
- You want an internal assistant whose answers can be checked against approved information and restricted by user access.
What the work involves
Prepare the document collection
Review source material, formats, ownership and version status. Identify duplicate or outdated content and prepare files for retrieval. Where scanned documents are included, assess the text recognition work needed and check its quality.
Build the retrieval pipeline
Organise documents into searchable passages, create the retrieval index and test it with representative questions. Preserve useful source details so users can locate the underlying document and understand the context of a cited passage.
Develop the assistant and access rules
Build an interface that answers from retrieved material and presents source citations. Configure access around the agreed user groups and document permissions. Define how the assistant responds when supporting information is absent or unclear.
Evaluate and maintain quality
Use questions prepared with subject specialists to assess retrieval relevance and answer accuracy. Examine missed information, unsupported statements and access behaviour. Document the process for adding, updating and removing sources after handover.
How the assignment runs
Select and prepare sources
Agree the document scope, intended users and common questions. Resolve source quality and access issues before building the initial collection.
Build and review answers
Implement retrieval and the assistant, then review results with people familiar with the subject matter.
Evaluate and hand over
Test accuracy, citations and access. Refine the system and document who maintains sources, reviews issues and approves changes.
What to prepare
These details will help us understand the starting point and agree a useful scope:
- A representative collection of documents with known owners and version status
- Common questions and examples of acceptable, source-supported answers
- User groups, access requirements and a plan for document updates
If some information is still being developed, we can discuss what is available in the first conversation.
Questions about this service.
For anything specific to your organisation, get in touch. We can discuss the requirements before you decide on an engagement.
Does a citation mean an answer is correct?
A citation lets someone inspect the supporting material. An answer can still misinterpret a passage or omit context, so we evaluate responses against representative questions. Users should review the source when acting on important information and report unsupported answers.
Can different teams have different document access?
Yes. Access requirements are part of the design. We map the intended user groups to permitted sources and test that retrieval respects those restrictions. Changes to document permissions also need a defined update process once the assistant is in use.
Can the service use PDFs and scanned records?
Typical sources include PDFs, Word documents, spreadsheets and web content. Scanned records may need optical character recognition first. The initial review checks representative files, because formatting, scan quality and document structure affect preparation and retrieval.
Let’s talk about what you need.
Share a little about your project, the challenge you are facing and where you would like some help.