04 · RAG & knowledge systems
Answers from your documents — with the receipts.
Retrieval-augmented generation (RAG) finds the passages that matter in your own content, then has a model answer from those passages only — and show where each answer came from.
Choose a question
Retrieved passages · relevance
- Access-Control-Standard.pdf §2.3Contractor accounts are limited to the systems named in their engagement…
- Records-Retention-Schedule.xlsxProcurement and tender records: retain for seven years after contract end…
- Travel-Policy.docx §5Domestic and international travel above $5,000 requires executive approval…
- Finance-System-Guide.pdf §1Access requests are lodged through the service desk and approved by the system owner…
- Onboarding-Checklist.docxNew starters receive email, intranet and timesheet access on day one…
How it works
Retrieve, rank, then answer.
Pick a question. The system scores every passage in the library for relevance, keeps the best ones, and writes an answer grounded in them — citing each source.
- Retrieve: hybrid search — meaning and keywords — over chunked, indexed documents.
- Rank: a re-ranking step keeps only the passages that genuinely answer the question.
- Answer: the model is instructed to use those passages only, and to say when they don’t contain the answer.
Sample documents and answers are invented for illustration.
Engineering the knowledge layer
Good answers start with good retrieval.
Document pipelines
Connectors for file shares, intranets, document management and databases, with parsing for PDFs, Office files and tables.
Chunking & embeddings
Structure-aware chunking, metadata and vector plus keyword indexes, refreshed as content changes.
Permission-aware search
Results filtered by the asker’s entitlements, mirrored from the source systems — so the AI never reveals what a person couldn’t open.
Citations
Every answer links to the passages used, so people can verify and go deeper.
Retrieval evaluation
Test questions with known sources, scoring whether the right passages are found and the answer is faithful to them.
Feedback loop
Unanswered and down-voted questions surface content gaps and tuning opportunities.
FAQ
RAG questions.
Why not just fine-tune a model on our documents?
Fine-tuning teaches style and patterns, not reliable facts, and it goes stale when documents change. Retrieval keeps answers current and traceable. We sometimes combine both.
What happens when the answer isn’t in our documents?
The system is instructed and tested to say so rather than guess, and those questions are logged so content owners can fill the gap.
Can it handle thousands of documents?
Yes. Retrieval scales with the index, not the model. Volume affects indexing and storage design, which we size during scoping.
Sitting on a library no one can search?
Tell us what content you have and who needs answers from it. We’ll propose a retrieval design and how we’d measure it.