RAG systems that cite their sources.
In short
A RAG (retrieval-augmented generation) system answers questions using your own documents rather than the model's memory: it retrieves the relevant passages first, then writes an answer grounded in them, with citations you can click through and check. That grounding is what separates a system you can put in front of staff from a chatbot that invents plausible answers. We build RAG systems on your content and ship them with an eval harness that scores retrieval and answer accuracy on a held-out set, so you know the number before you launch, not after.
What you get
What changes when this works.
Answers with a paper trail
Every response points at the passage it came from. A reviewer can verify an answer in seconds instead of re-reading the source.
A measured accuracy score
We build the eval set alongside the system. You get retrieval and answer-quality numbers on held-out questions, before launch and continuously after.
Honest failure
The system says “I don’t know” when the documents don’t support an answer. Confident wrong answers are the failure mode that destroys internal trust, so we tune against them explicitly.
Content that stays current
An ingestion pipeline that handles updates, deletions and permissions, so the index doesn’t quietly drift away from the source of truth.
Concretely
What we actually hand over.
No deliverable on this list is a document about the work. Each one is the work.
- Document ingestion and chunking pipeline tuned to your file types
- Vector store (pgvector, Qdrant, Pinecone or Weaviate, chosen on your constraints rather than our habits)
- Retrieval layer with hybrid search and reranking where it earns its cost
- Answer generation with inline citations and refusal behaviour
- Eval harness with a labelled question set and CI-runnable scoring
- Permissions model so retrieval respects who is allowed to see what
In practice
A European insurance carrier
Our clients’ systems are commercially sensitive, so these stories are anonymised at their request. Metrics are reported as ranges or percentages rather than raw figures.
Problem
Claims analysts answered policy questions by opening and re-reading long wordings by hand, and answers varied between analysts.
Action
We built a RAG layer over the policy library with clause-level citations, plus an eval set of real analyst questions graded against approved answers.
Result
[X]% of questions answered correctly on the held-out set, and answer time cut from [X] minutes to under [X].
“The citations are the reason it got approved. Our reviewers can check the system’s working rather than trust it.”
Questions
RAG systems, answered plainly.
What is a RAG system?
A RAG system retrieves relevant passages from your own documents and uses them to write a grounded, cited answer, instead of relying on what a language model memorised during training. The retrieval step is what makes the answer checkable: you can see the source passage behind every claim.
What’s the difference between RAG and fine-tuning?
RAG changes what the model knows; fine-tuning changes how the model behaves. If your problem is “the model doesn’t know our policies”, that is a RAG problem, and RAG updates the moment a document changes. If your problem is “the model won’t follow our format or tone”, that is a fine-tuning problem. Most business problems we see are retrieval problems, and RAG is far cheaper to keep current.
Will it hallucinate?
Any language model can generate an unsupported claim. What changes the risk is grounding and measurement: answers cite the passages they came from, the system is tuned to refuse rather than guess when the documents don't support an answer, and the eval harness scores exactly that behaviour on held-out questions so you can see the rate rather than hope about it.
Can it work with our permissions?
Yes, and it has to. Retrieval is filtered by the requesting user’s access before any passage reaches the model, so a system that indexes everything still only answers from what that person is entitled to see.
How much of our data leaves our environment?
That is a scoping decision, and we make it explicitly on Day 0. We build against hosted APIs, cloud-provider models such as AWS Bedrock, or fully self-hosted models depending on your constraints, and we will tell you plainly what each option costs in accuracy and money.
Related
The other four.
Most projects touch more than one of these. If you're not sure which yours is, that's what the scope call is for.
Start here
Scope your first workflow.
Tell us the process that eats your team’s afternoons. We’ll come back within one business day with a scope, a fixed price, and an honest answer about whether AI is the right tool for it.
- No sales sequence. One reply, from the person who’d build it.
- If we don’t think it’s worth building, we’ll say so.
Email us what the process is, roughly how often it runs, and who does it today. That is enough for us to come back with a scope.
Email usOr write to admin@vintrexlabs.com directly.
Prefer to talk
Book a 20-minute call.
Bring one workflow. We’ll tell you on the call whether it fits in 10 days, and roughly what it costs.
Pick a 20-minute slot
Opens our live calendar. Bring one workflow and we’ll tell you on the call whether it fits in ten days.
Loads cal.com in this page. No account needed.