Support copilot
Answers customer questions from your help centre and ticket history, hands off to an agent with a draft and sources when confidence is low.
Deflect and assist
AI Knowledge Base
AI knowledge bases that answer from your documents
We design, build and run retrieval-augmented generation (RAG) systems that give staff and customers cited, permission-aware answers from content you already own.
- Every answer cites its sources
- Permissions enforced at retrieval
- Claude, GPT, Gemini or open-weight models
Who this is for
When a knowledge base pays for itself
A RAG project makes sense when people spend real time hunting for answers that already exist somewhere: a wiki, a ticket history, a contract folder, a product manual.
The typical buyer is a support lead with rising ticket volume, an operations head whose team keeps re-asking the same questions, or a CTO who has been told to "add AI" and wants something that holds up in production. The common thread is content that is valuable, scattered and slow to search.
We have built knowledge bots into our own event platform, Eventify, and we build them for clients across support, operations, legal and HR. The work is the same each time: get the content right, make retrieval precise, prove the answers are correct, then keep it that way.
Answers you can check
Every response cites the chunk it came from, with a link back to the page, section or ticket.
Permissions carried through
Access labels travel with each chunk and are applied inside the retrieval query, never after the fact.
Measured, not assumed
A labelled test set and a scoreboard: accuracy, citation rate and refusals, tracked on every release.
Fresh within hours
Connectors re-index changed content on a schedule you set, so answers do not go stale when policies change.
Use cases
Three shapes a knowledge base usually takes
Same engine, different front door. Most clients start with one of these and add the others once retrieval quality is proven.
Internal copilot
One place for staff to ask about policies, procedures, projects and systems, with answers filtered by what each person is allowed to see.
Permissions-aware answersDocument Q&A
Ask questions of long contracts, manuals, policies and reports, and get page-level citations rather than a summary you cannot verify.
Question your documents
Hard truths
Why RAG projects stall
The demo is easy. A weekend with a framework produces something that answers nicely on ten hand-picked questions. Production is where projects stop, and the causes repeat.
The six patterns we see most often
- Content nobody owns. Outdated, duplicated or contradictory pages are retrieved just as readily as good ones. The model faithfully cites the wrong answer.
- No evaluation set. Without 100 to 300 labelled questions, every change is a guess and nobody can say whether the system got better.
- Chunking by default. Fixed 1,000-character slices cut tables in half and separate headings from the text they explain.
- Permissions bolted on later. Filtering results after retrieval leaks existence of documents and breaks top-k quality.
- Pure vector search. Product codes, people's names and error strings need keyword matching too. Hybrid search fixes this.
- No feedback loop. Users stop reporting bad answers after the second week unless reporting is one click and visibly acted upon.
Each of these is a design decision, not a research problem. Our process below is built to force those decisions early, while they are still cheap.
How we work
From content audit to a system you can trust
Five stages, each with a deliverable you can inspect. A pilot on a single source typically runs four to six weeks; a multi-source production build eight to sixteen.
Audit the content
We inventory sources, owners, formats and freshness, and write the first 100 evaluation questions with your subject-matter experts. Output: content map and test set.
Design retrieval
Chunking rules per document type, metadata and access labels, embedding model, hybrid search and reranking, vector store. Output: retrieval design and a cost model.
Build and integrate
Connectors, ingestion pipeline, retrieval service, prompt and citation logic, the chat or search interface, and single sign-on. Output: working system on your data.
Evaluate and harden
Scores on the test set, prompt-injection testing, confidence gates and refusal behaviour, following our secure RAG architecture. Output: evaluation report and go-live checklist.
Operate and improve
Monitoring, re-indexing, monthly review of misses and escalations, and model or index upgrades when they earn their keep. Output: a scoreboard that moves in the right direction.
Deciding
RAG, fine-tuning or a bigger context window?
Clients often arrive with one of these in mind. They solve different problems, and most production knowledge bases use RAG first and the others only when measured need appears.
| Question | RAG (retrieval) | Fine-tuning | Long context only |
|---|---|---|---|
| Content changes weekly | Re-index, no retraining | Retrain each time | Re-send each time |
| Needs citations | Native | No provenance | Possible, but weak |
| Per-user permissions | Filter at retrieval | Model knows everything | Must pre-filter manually |
| Changing tone or format | Prompting usually enough | Strongest option | Prompting usually enough |
| Corpus size | Millions of pages | Limited by training data | A few hundred pages per call |
| Cost per question | Low to moderate | Low at run time, high up front | High, grows with corpus |
Our guide to prompt engineering vs fine-tuning goes deeper, and fine-tuning and evals is a service in its own right.
Services in this section
AI knowledge base services
Enterprise RAG
Reference architecture, connectors for SharePoint, Google Drive, Confluence, Zendesk, Salesforce and databases, security layers and evals.
ExploreDocument Q&A
Question long, structured documents with page-level citations: contracts, manuals, policies, reports.
ExploreSupport copilot
Customer-facing deflection and agent assist from one knowledge base, with a confidence gate and clean hand-off.
ExploreInternal copilot
Staff Q&A across policies, systems and projects, with permissions enforced inside the retrieval query.
ExploreKnowledge base migration
Turn a legacy wiki or help centre into AI-ready content: audit, structure, metadata, embeddings, validation.
ExploreVector search
The retrieval layer underneath: semantic and hybrid search, embeddings strategy and vector database selection.
Explore
Related
Read before you build
Guide
RAG explained
How grounded answers work, with the trade-offs at each stage.
Read ServiceSecure RAG architecture
Injection defence, data boundaries and output checks for retrieval systems.
Read SolutionKnowledge management
The organisational side: ownership, review cycles and adoption.
Read ServiceLLM app development
When the knowledge base is one part of a larger AI application.
Read
FAQ
Frequently asked questions
What is an AI knowledge base?
A system that answers questions from your own content rather than from a model's general training. Documents are split into chunks, converted to embeddings and stored in a vector index. At question time the most relevant chunks are retrieved and handed to a large language model, which writes an answer with citations. The technique is called retrieval-augmented generation; our guide RAG explained walks through it step by step.
How long does a RAG project take?
A focused pilot on one content source typically takes four to six weeks, including a labelled evaluation set. A production system with several connectors, permissions and monitoring usually takes eight to sixteen weeks. Migrating a large legacy knowledge base adds time for the content audit; see knowledge base migration.
Which LLM and vector database do you use?
Whichever fits your constraints. We work with Anthropic Claude, OpenAI, Google Gemini and open-weight models, and with PostgreSQL + pgvector, Milvus, Pinecone, Qdrant and Weaviate. We run Milvus and pgvector in our own production systems. The vector database chooser gives a first recommendation in two minutes.
Can the knowledge base respect our existing permissions?
Yes, and it must. We store access labels on every chunk and apply them as a filter inside the retrieval query, so a user never receives a chunk they could not open in the source system. The approach is described on the internal copilot page and in our secure RAG architecture service.
How do you stop the model making things up?
Grounding, constraints and measurement. Answers must cite retrieved chunks, the prompt instructs the model to decline when evidence is missing, and a confidence gate routes weak answers to a human. We measure citation rate and factual accuracy on a labelled test set before and after every change. Our evals service covers the method.
What does it cost to run?
Three cost lines matter: one-off indexing, per-query LLM usage and vector storage. The RAG cost estimator shows all three with editable unit prices, so you can model your own provider and volumes before committing.