AI Services
RAG Knowledge Base
Retrieval-augmented generation, in plain terms: find the right document first, then answer from it. The finding is where these systems succeed or fail.
A RAG knowledge base answers questions from your own documents and shows which one it used. The Nexclick builds the retrieval layer properly, because most poor AI answers are retrieval failures rather than model failures — the system found the wrong document, not the wrong words.
Is this you?
What usually prompts the call
- Information exists across hundreds of documents and finding the right answer takes half an hour.
- New staff take months to become useful because the knowledge is in people rather than documents.
- You have technical specifications, policies or contracts nobody can search meaningfully.
- You tried a generic AI tool on your documents and the answers were plausible and wrong.
What we do
The actual deliverables
Things that appear on an invoice, not adjectives.
- Content audit and preparation
- What exists, what is current, and what contradicts itself. Feeding a model three versions of the same policy produces confident answers from the wrong one.
- Chunking and indexing strategy
- How documents are split determines what can be retrieved. Splitting a contract mid-clause or a manual mid-procedure is the most common cause of poor answers.
- Retrieval quality testing
- An evaluation set of real questions with known correct sources, so retrieval is measured rather than assumed. This is the step that separates working systems from demos.
- Citation and traceability
- Every answer showing its source document and section, so a person can verify it. Non-negotiable for anything used in a regulated or technical context.
- Permission-aware retrieval
- People only get answers from documents they are entitled to see. A knowledge base that leaks HR or commercial documents across departments is a serious problem.
- Refresh and maintenance
- How new and updated documents enter the index. A knowledge base that silently goes stale becomes actively misleading.
- Interface appropriate to the users
- Search, chat, or embedded in the tool people already use. Adoption depends more on where it lives than on how good the answers are.
Decision tree
Why is your AI giving wrong answers?
Almost every complaint about AI answer quality traces to one of these, and only the last is about the model. Work down in order — the fixes are very different.
01The correct information is not in the documents at all
A documentation problem, not an AI one. No retrieval system invents institutional knowledge that was never written down.
02Several documents say different things
Content conflict. The system retrieves one of them confidently. Resolve the contradiction in the source material — nothing downstream fixes it.
03It retrieves a related but wrong document
Retrieval tuning. Usually chunking or embedding strategy. This is the most common technical cause and it is fixable.
04It answers from half a procedure or half a clause
Chunking problem. Documents split mid-section lose the context that made them meaningful. Restructure how they are divided.
05It answers from a superseded version of a document
Refresh and versioning. The index contains material that should have been removed. Fix the maintenance process.
06It answers correctly but too vaguely to act on
Prompt and output formatting, not retrieval. Ask for specifics, cite sections, and constrain the response shape.
07It answers confidently about something outside the documents entirely
Scope boundaries missing. The system should decline rather than fall back on model knowledge — that is a configuration decision.
How it works
Step by step, with timeframes
Timeframes are typical rather than guaranteed, and they assume we get account access and approvals when we ask.
- 01Week 1–2
Audit the content
What exists, what is current, what conflicts. Often surfaces that the documentation problem is bigger than the AI problem.
- 02Week 2–3
Build an evaluation set
Real questions with known correct sources, agreed before building, so quality can be measured rather than debated.
- 03Week 3–7
Build and tune retrieval
Chunking, indexing and retrieval tested against the evaluation set and adjusted until accuracy is acceptable.
- 04Week 7–10
Interface, permissions and launch
Access controls, interface and refresh process, then a pilot group before wider rollout.
What you get
Reporting and ownership
- An evaluation set of real questions with known answers, so accuracy is measured not claimed.
- Retrieval accuracy reported as a figure before launch, and re-measured after changes.
- Citations on every answer, traceable to document and section.
- Permission-aware access, so nobody sees answers from documents they should not.
- A documented refresh process, so the index does not silently go stale.
Tools and platforms
- Commercial LLM APIs
- Vector databases (Pinecone, pgvector, Qdrant)
- Document parsing and chunking pipelines
- Evaluation frameworks
- Existing document stores — SharePoint, Drive, Confluence
Timeline
How long this actually takes
Seven to ten weeks. The content audit frequently reveals that the harder problem is documentation rather than AI — contradictory policies, superseded versions, and knowledge that was never written down. Where that is the case we will say so, because no retrieval system resolves a contradiction it was given. Expect accuracy to be good rather than perfect: a well-built system retrieves the right source most of the time, and the citation is what makes the remainder safe rather than dangerous.
Pricing model
Project, then retainer
Fixed price for the build, then a small monthly fee for hosting, refresh and monitoring. Model and vector database costs are yours directly and shown at cost.
Questions
RAG Knowledge Base questions
What does RAG actually mean?
Retrieval-augmented generation. The system searches your documents for relevant passages, then asks a language model to answer using only those passages. The retrieval half is where quality is won or lost, and it is the half most implementations underinvest in.
Why not just upload documents to ChatGPT?
Fine for a handful of documents and a single user. It does not scale to hundreds of documents, does not respect your permission structure, does not stay current as documents change, and does not give you a way to measure whether answers are right.
How accurate will it be?
Measurably, which is the important part. We build an evaluation set of real questions with known correct sources and report retrieval accuracy as a figure before launch. Good systems retrieve correctly most of the time, and the citation is what makes the rest safe.
Can it handle scanned documents and PDFs?
Yes, with an extraction step — that is where document processing and OCR overlaps this service. Scanned quality varies and poor scans produce poor extraction, which produces poor retrieval. It gets assessed during the content audit rather than assumed.
Will people see documents they should not?
Not with permission-aware retrieval, which filters by the user’s entitlements before anything is retrieved. This is essential in any organisation with HR, commercial or client-confidential material, and it needs designing in rather than adding later.
What happens when documents change?
A refresh process re-indexes new and updated material, and removes superseded versions. Without it, the system quietly answers from last year’s policy — which is worse than having no system, because people trust it.
Last reviewed 28 July 2026.
Tell us what you are trying to fix
A 20-minute call, no pitch deck. The Nexclick will tell you what we would do, roughly what it costs, and whether we are the right people for it.