Source concept
Which documents and systems are included and how they are prepared and kept current.
AI & automation
RAG solutions for enterprise knowledge
Turn enterprise knowledge into answers with sources, permissions and measured quality.
In short
Ki4data builds RAG solutions that answer questions from your own documents and cite the source of every statement. We select suitable sources, build a prototype with permission checks and measure answer quality against agreed test questions before you decide on production use. This makes sense when knowledge sits in policies, manuals or technical documentation and a lot of time goes into searching.
What is RAG?
Retrieval-augmented generation (RAG) combines searching for relevant content with generating an answer based on that content; the approach was described by Lewis et al. in 2020. The language model answers from the passages found in your documents rather than from its general training, and can cite them as sources.
Last updated:
Typical starting points
- Internal policies, manuals or product documentation are extensive and hard to search.
- The same questions keep landing with the same experts.
- A first chatbot attempt gave answers without evidence or with outdated information.
- Document permissions must also be enforced inside an AI application.
What you receive
Working prototype
A usable application on a defined knowledge area, including source citations.
Test questions
A question set with expected answers, agreed with the business.
Evaluation report
Measured answer quality, source grounding, latency and cost, with documented limits.
Integration and operating plan
How the solution is connected, monitored and developed further.
How we work
Sources
Select documents and systems and assess their quality.
Preparation and index
Split, enrich and index the content.
Permissions
Carry source-system permissions into the search.
Retrieval
Find and rank the passages relevant to a question.
Model and answer
Generate the answer from those passages and cite sources.
Evaluation
Measure with test questions, analyse errors and improve deliberately.
How we measure quality
- Source coverage: are the relevant documents included?
- Permission fidelity: does everyone see only what they are allowed to?
- Freshness: are source changes picked up promptly?
- Grounding: can every statement be traced to a source?
- Missing information: does the solution say when it has no answer?
- Latency and cost per request
Prerequisites and limits
A good fit when …
- Knowledge exists as text and changes regularly.
- Answers must be traceable and backed by a source.
- There is a clearly bounded first knowledge area and a business contact.
Another solution fits better when …
- People mainly want to find a specific document; good classic search is often faster and cheaper.
- Answers should come from structured metrics; an approach on the data warehouse fits better.
What drives effort and cost
Every project is scoped individually. These factors influence the effort most:
- Number, format and quality of sources
- Permission and integration requirements
- Scope of evaluation and test questions
- Usage volume and chosen model in production
Frequently asked questions
When is classic search enough?
When users mainly want to find one document and read the answer in it themselves. RAG pays off when answers have to be assembled from several passages or questions are asked in natural language.
How are permissions enforced?
Source-system permissions are carried along at indexing time and checked on every request before any passage reaches the model. The answer therefore cannot contain content the person asking is not allowed to see.
How do we catch wrong answers?
With an agreed test question set that is rerun before every change, and by evaluating user feedback in production. Answers without sufficient sources are flagged as such instead of guessed.
RAG or fine-tuning?
RAG fits when answers depend on current, changing knowledge and must be verifiable: new documents are indexed rather than retrained, and permissions can be checked per request. Fine-tuning fits better when a model should learn a particular style or fixed task format. The two can be combined.
Related services
AI agents for clearly defined business processesAgents with defined tools, execution limits and traceable logging.Data warehouse consulting and deliveryBuild or modernize one reliable data foundation for reporting, analytics and AI.Operating and optimizing your data & AI solutionsOperations, maintenance, monitoring, support and performance optimization for data & AI solutions.
Next step
In a first conversation we clarify your situation, goals and a sensible starting point. Afterwards you receive an assessment of what a project could look like.
Assess your RAG use case