RAG systems and AI knowledge bases
Retrieval-augmented generation lets a model answer using your documents instead of its training data. Done well it is the single highest-value AI pattern for most businesses. Done badly it produces confident nonsense with a citation attached.
What this involves
Ninety per cent of RAG quality is decided before the model is involved. How documents are split, what metadata travels with each chunk, how the query is rewritten, and whether results are reranked before they reach the model — those choices determine whether answers are right.
We have seen far more RAG systems fail on chunking than on model choice. A table split across two chunks, or a policy document where the exception clause is separated from the rule, produces answers that are wrong in ways nobody notices until it matters.
- Document ingestion from your storage, CMS, wiki or database
- Chunking strategies chosen per document type, not one size fits all
- Vector search with metadata filtering and permissions
- Reranking to push the right passage to the top
- Citations linking back to the exact source passage
- An evaluation set that proves accuracy and catches regressions
Permissions are part of retrieval
If your documents have access rules, the retrieval layer must enforce them — not the prompt. A system that retrieves a restricted HR document and then politely declines to discuss it has already leaked it into a log. Access filtering happens at the query, before anything reaches the model.
Keeping it current
Knowledge bases rot. We build the ingestion as a repeatable pipeline with change detection, so updating a source document updates the index within minutes, and removing one removes it from answers. Manual re-indexing is how a knowledge base quietly becomes wrong.
Proving it works
Before launch we agree a test set of real questions with known correct answers. The system is measured on retrieval accuracy and answer quality separately, because they fail for different reasons and need different fixes. You get the numbers, not an assurance.
Frequently asked questions
How is RAG different from fine-tuning?
Retrieval gives a model facts at question time; fine-tuning changes how a model behaves. If you want correct, current, citable answers about your business, retrieval is almost always the right tool and it is cheaper to keep up to date. We have written a longer comparison in the insights section.
How many documents can it handle?
Millions of chunks is routine with modern vector databases. Volume is rarely the constraint — document quality is. Ten thousand pages of contradictory, undated internal wiki pages is a much harder problem than a million clean product records.
Where is the data stored?
Wherever you require. Your cloud account, your region, self-hosted vector database if needed. For regulated clients we keep the whole pipeline inside their infrastructure with no data crossing borders.
What if the answer is not in our documents?
It says so. That behaviour is enforced in the prompt and verified in the evaluation set — a system that never admits ignorance is a system that will eventually invent something important.
Tell us what you are building.
Send a short description of the problem and we will reply within one business day with an honest view of scope, cost and whether we are the right person for it.
Or email directly: contact@hire-ai-dev.com