The knowledge problem
Every organisation has the answers somewhere — in a policy PDF, a Confluence page, a closed ticket, a spreadsheet on someone’s drive. Finding them is the problem. New staff ask colleagues; colleagues get interrupted; customers wait. Traditional enterprise search returns ten links when people wanted one sentence.
A well-built RAG assistant returns the sentence — and the link to where it came from.
Retrieval is where projects succeed or fail
Most disappointing RAG systems have the same root cause: the right passage was never retrieved, so the model answered from nothing. We treat retrieval as its own engineering problem and measure it separately:
- Parsing that respects structure. Tables stay tables, headings stay attached to their sections, scanned pages get OCR.
- Chunking per document type. Policies, product manuals and support tickets need different strategies.
- Hybrid search. Vector similarity finds meaning; keyword search finds exact product codes and names. We use both, then re-rank.
- Metadata filters. Region, product, date and permissions narrow the search before the model sees anything.
Only once the right evidence is reliably found do we tune how answers are written.
Grounded, cited, honest
Answers include citations to the exact passages used. When evidence is missing or contradictory, the assistant says so and points to the closest sources or a human contact. Conversation memory is handled carefully so follow-up questions work without leaking context between users.
Measured before and after launch
We build an evaluation set from real questions — not invented ones — and score retrieval recall, answer correctness and faithfulness. After launch, thumbs-up/down feedback and unanswered questions feed a weekly review that usually reveals content gaps as much as technical ones. Fixing a missing policy page often helps more than changing a model.