RAG System Development
Answers from your own documents, with the source attached every time.
What is RAG systems?
A RAG system finds the relevant passages in your own documents and gives them to a language model to answer from, so responses are grounded in your material and can be traced back to the exact source.
Retrieval is how you make a model know your business, and it is what most people mean when they ask about fine-tuning. Your documents are searched, the relevant parts are put in front of the model, and it answers from those rather than from general knowledge.
Two properties make it worth building. The answer cites its source, so anybody can check it. And when nothing relevant is found, the correct response is to say so rather than to produce something plausible.
That second property is where most implementations fail. A system that always answers is easy to build and quietly dangerous, because a confident answer assembled from irrelevant passages looks exactly like a good one.
Why we build the refusal first
Duara builds retrieval systems from Nairobi over policy libraries, technical documentation, contracts, procedures and support histories.
We build the failure to answer before the answer. A system that says it does not know is more useful than one that always produces something, and it is harder to build because it requires deciding what counts as insufficient evidence.
Who this is for
- Staff cannot find answers in your own documents
- The same questions are asked repeatedly
- Documents are numerous and rarely read
- You want AI that knows your material specifically
Industries we serve
- Financial services
- Legal and professional services
- Healthcare providers
- Engineering and technical services
- Insurance
- Large public sector suppliers
Problems this solves
These are retrieval problems. The information exists and cannot be reached at the moment somebody needs it.
- Answers exist and cannot be found. The policy covers the situation and locating it takes twenty minutes, so people ask a colleague instead.
- The same questions repeatedly. Experienced staff spend their time answering questions the documentation already covers.
- Documents nobody reads. A large library exists, is maintained, and is consulted far less than the effort of maintaining it would justify.
- Generic AI that does not know you. A general model answers plausibly about your policies and is simply making it up.
What we deliver
Your documents indexed and searched properly, answers generated only from retrieved passages, sources cited on every response, and an explicit refusal where the material does not cover the question.
Your documents indexed and searchable
Answers generated only from retrieved material
A source citation on every answer
An explicit refusal where evidence is insufficient
Permissions respected so people see only their material
Reporting on questions that could not be answered
Features
- It says when it does not know. Insufficient evidence produces a refusal rather than a plausible assembly, which is the hardest part to build properly.
- Every answer cites its source. The passage and document are shown, so an answer can be verified rather than trusted.
- Permissions are respected. Retrieval honours your access rules, so nobody receives an answer drawn from a document they could not open.
- Unanswered questions reported. What people asked and the system could not answer is usually a gap in your documentation worth knowing about.
- Document currency surfaced. Where two documents conflict or one is clearly out of date, that is shown rather than silently resolved.
Benefits
- Answers arrive in seconds. The information you already maintain becomes reachable at the moment somebody needs it.
- Experienced staff get their time back. Repeated questions stop consuming the people who know the answers.
- Answers can be checked. A cited source means a wrong answer is detectable rather than merely convincing.
- Documentation gaps become visible. The unanswered question report tells you what your material does not cover.
Technologies we use
- Document ingestion. Processing your material into searchable form, including handling of formats and structure.
- Chunking and indexing. Splitting documents so retrieved passages carry enough context to be useful.
- Semantic and keyword retrieval. Combining meaning based and exact matching, since each fails on cases the other handles.
- Grounding and citation. Constraining answers to retrieved material and attaching the source.
- Sufficiency checking. Deciding when retrieved evidence is too weak to answer, and refusing instead.
Retrieval quality decides everything. A perfect model answering from the wrong three passages gives a wrong answer confidently, and no amount of prompt work fixes bad retrieval.
How the process works, step by step
A retrieval system usually takes five to nine weeks, and it starts by testing retrieval quality before any answering is built.
-
Step 1
Free document review
We look at what material exists, its condition and whether it actually contains the answers people need.
-
Step 2
Retrieval tested first
We measure whether the right passages come back before building any answering, since nothing works if that fails.
-
Step 3
Refusal behaviour built
The system learns to say it does not know before it learns to answer, which is the correct order.
-
Step 4
Citations and permissions
Sources are attached and access rules applied so answers respect who is asking.
-
Step 5
Live with gap reporting
Unanswered questions are collected, which usually improves your documentation as much as the system.
Quality assurance and how we work
The commitments attached to a system that answers from your own material.
- We build the refusal before the answer. A system that always produces something is easier to build and quietly dangerous, since a confident wrong answer looks like a good one.
- Every answer carries its source. Without a citation an answer can only be trusted, and trusted is not the same as correct.
- Retrieval quality is tested before answering is built. A perfect model answering from the wrong passages is wrong confidently, and no prompt work fixes that.
- Access permissions are enforced in retrieval. Nobody receives an answer drawn from a document they could not open themselves.
- We report what it could not answer. Those questions are usually a documentation gap, and hiding them would waste the most useful output.
Frequently asked questions about RAG systems
Is this what we actually need instead of fine-tuning?
Usually yes. If you want a model to know your material, retrieval is the right technique and it is cheaper and updateable.
Will it make things up?
It answers from retrieved passages and refuses where evidence is insufficient. That refusal is built first, deliberately.
Why build the refusal first?
Because a system that always answers is easy to build and quietly dangerous. A confident answer from irrelevant passages looks exactly like a good one.
Can we see where an answer came from?
Yes, the passage and document are cited every time, so an answer can be verified rather than trusted.
What if two of our policies disagree?
The conflict is shown rather than silently resolved, and it is frequently the most useful thing the system surfaces.
Does it respect who can see what?
Yes. Retrieval enforces your access rules, so nobody gets an answer from a document they could not open.
What happens when it cannot answer?
It says so, and the question is logged. Those logs usually point at a genuine gap in your documentation.
How long before it is answering questions?
Five to nine weeks, with retrieval quality tested before any answering is built.
Why test retrieval separately?
Because a perfect model answering from the wrong three passages is wrong confidently, and no amount of prompt work fixes bad retrieval.
What document formats can it use?
Most common ones. Scanned documents need OCR first, and quality there sets the ceiling.
How quickly do document updates take effect?
As current as your documents. When a document is updated the system reflects it, which is the main advantage over fine-tuning.
Is there a limit on library size?
Large libraries are normal. Volume matters less than whether the material actually contains the answers people need.
Can it work over emails and support history?
Yes, and past support conversations are often a better source than formal documentation for practical questions.
Can it answer from Swahili documents?
Yes, including retrieving from documents in one language and answering in another, though we test that on your material.
Can staff correct a wrong answer?
Yes, and corrections usually point at a retrieval or documentation problem worth fixing rather than a model failure.
Do our documents leave our systems?
No. Indexing and retrieval run in your infrastructure, and what is sent to a model is scoped deliberately.
Will it replace our knowledge team?
No. It makes their material reachable, and the gap reporting usually gives them a clearer improvement list than they had.
Can it answer customers directly?
Where the material supports it, though we would insist on the refusal path and an escalation route to a person.
Who can see what people asked?
You do, under your own permissions. Question logs stay in your systems and train no public model.
What does a retrieval system cost?
Document volume and query rate determine it, quoted after the document review.
Glossary of terms
Retrieval terms in plain language.
- RAG
- Retrieval augmented generation, meaning answers generated from documents retrieved at the time of asking.
- Chunking
- Splitting documents into passages small enough to retrieve and large enough to make sense.
- Grounding
- Constraining an answer to the retrieved material rather than general model knowledge.
- Semantic search
- Matching on meaning rather than exact words, which finds relevant passages worded differently.
- Sufficiency
- Whether the retrieved evidence is strong enough to answer at all.
- Citation
- The pointer from an answer back to the passage it came from.
Related services in AI Development
Talk to us about RAG systems
Pick a question your staff ask constantly that your documents already answer. That one shows whether retrieval is worth building.
Start your project