A general model was trained on the public web up to a date. Your policy, your prices, last month's decision are not in it, and asked about them it answers anyway.
The model was trained on the public web up to a point. Your prices, your policy, last month's decision: none of it is there.
Asked about something it never saw, a model produces a fluent answer anyway. In a support queue or a contract review that is a liability.
Without a reference, nobody can check the answer. The person who reads it has to trust it or redo the work.
A RAG system looks up the relevant passages in your knowledge base before the model writes a word. The answer is grounded in your documents, current as of the last index, and carries its citations.
Three parts, each one measurable on its own before the next is trusted.
Your material becomes a searchable base.
The right passages, for the question asked.
A grounded answer with its sources attached.
Four phases, each with a deliverable you can test.
What people ask, where the answers live, who may see what.
Ingest, split, embed and index your documents.
Wire the search into the assistant your teams use.
Ship, watch, and keep the base current.
Answers you can check, from the knowledge you already have.
The model writes from your passages, so an answer with no source in the base is refused rather than guessed.
The person asking gets the passage and the answer together, without opening four systems.
A citation on every response means the reader can verify it in one click, and does.
| Side by side | Model alone | With RAG |
|---|---|---|
| Where answers come from | Training data, fixed at a date | Your documents, current at the last index |
| Company knowledge | None | Everything you index, by permission |
| Sources | None | Document and section on every answer |
| When it does not know | Guesses | Says so |
| Control | None over what it knows | You decide what is indexed and who sees it |
Anything with text: PDFs, wikis, tickets, emails, database records, transcripts. Each source keeps its owner and access level, so the assistant only retrieves what the asker may see.
Sources are re-indexed on a schedule you set, from hourly to weekly, and on demand. An answer always states the date of the passage it used.
Yes, with permission filtering at retrieval time: a passage is only considered if the person asking may read the source. The model can run locally when the documents may not leave your perimeter.
Against a question set agreed in discovery, with the expected answers and sources. We report recall, citation accuracy and refusal rate, and review a weekly sample of live answers with your team.
We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.
