RAG & knowledge retrieval
Answers grounded in your own documents, with a citation on every claim, your existing permissions enforced at retrieval, and an honest "I don't know" when the answer is not in the corpus.
Retrieval is
the whole game.
Retrieval-augmented generation is often described as "point a model at your documents." That framing hides where all the difficulty lives. If the right passage is not in the context window, no model can answer correctly — it can only answer confidently.
So the engineering is almost entirely upstream of the model: how documents are parsed, how they are split, how tables and headers survive that split, how a query is expanded, how candidates are re-ranked, and what happens when the top result is merely topical rather than actually responsive.
The demo always works, because the demo asks a question whose answer is on page one of a document someone chose. Production asks about a 2019 amendment buried in an appendix that contradicts the main policy.
What we actually build
Seven stages. The model appears in exactly one of them.
Ingestion & parsing
Your documents as they exist: PDFs, scans, Word files, wiki pages, ticket histories, spreadsheets. Layout-aware parsing so a table stays a table and a heading stays attached to the text it governs.
Chunking with structure
Split on document structure rather than a fixed character count, with section context carried into every chunk — so a clause retrieved on its own still knows which policy and which version it belongs to.
Hybrid retrieval
Semantic search combined with keyword matching. Pure vector search reliably misses exact identifiers — part numbers, statute references, policy codes — which is precisely what people search for.
Re-ranking
A second pass that scores candidates against the actual question rather than general similarity. This is usually the single largest quality gain available, and it is routinely skipped.
Permission-aware filtering
Access control applied at retrieval, before generation. A user's answer can only be built from documents that user is entitled to read — enforced server-side, never in the prompt.
Grounded generation
Every claim carries a citation to a specific passage the reader can open. When the retrieved context does not contain an answer, the system says so instead of composing a plausible one.
Freshness
Incremental re-indexing when source documents change, with a visible indicator of how current the index is — because a confident answer from a superseded policy is worse than no answer.
Measured on questions, not vibes
Before we build, your subject-matter experts write a question set — typically 80 to 150 real questions — and, for each, the document and passage that actually contains the answer. Including the questions whose answer is nowhere, because refusing correctly is a measurable skill.
We then measure two things separately: retrieval (was the right passage in the top results?) and answer quality (did the response use it correctly, and is the citation real?). Reporting one blended score hides which half is broken, and they are fixed by completely different work.
Why a self-graded score isn't accuracyThe citation is the product
For internal knowledge work, a fluent answer without a source is close to worthless — the reader still has to verify it, which was the original job.
So every answer links to the passage it came from, and we test that the citation genuinely supports the claim rather than merely appearing near it. A fabricated or mismatched citation counts as a failure in the evaluation, not a formatting issue.
Fixed price. Written exclusions.
Prices in Canadian dollars, verified 23 August 2026. An advisory hour is CAD $195; fixed-scope advisory engagements start at CAD $4,500; minimum build engagement is CAD $9,500.
Departmental assistant
from $24,000CAD · One corpus, one access tier, web interfaceIncluded
- All seven stages above, in production
- Labelled question set with agreed thresholds
- Separate retrieval and answer-quality reporting
- Citation-linked answer interface
- Runbook and full source handover
Not included
- Per-user permission filtering
- Ongoing hosting — see run tier
Multi-source & permissioned
$52,000CAD · Several systems, per-user access control, Teams or SlackIncluded
- Everything in the departmental tier
- Retrieval across several connected systems
- Per-user permission filtering at retrieval
- Delivery inside Teams, Slack or your own app
- Query logging and unanswered-question reporting
Not included
- Cleaning up or reconciling contradictory source documents
- Building a permission model you don't have
Hosting & care
from $2,200CAD / month · Or run it in your own tenancyIncluded
- Hosting in a Canadian or U.S. region
- Continuous re-indexing as documents change
- Monthly question-set re-run with a trend report
- Retrieval tuning from real unanswered questions
Not included
- LLM and embedding inference — at cost
- New source systems
When you should not buy this
When your documents contradict each other. Retrieval will faithfully surface both the 2019 policy and the 2024 one that replaced it, and the system has no principled way to know which governs. If your corpus has no versioning and no owner, fixing that comes first — and it is genuinely more valuable than this project.
When the knowledge isn't written down. If the real answers live in people's heads and the documents are outdated, you would be building a very sophisticated way to retrieve wrong information.
When search would do. If people mostly need to find the document rather than synthesise an answer across several, good search is cheaper, faster and easier to trust. Some of the best outcomes here are a better search index and no model at all.
When the corpus is small enough to fit in a prompt. Under a few hundred pages, sending the relevant documents directly to a long-context model is simpler and often more accurate. Retrieval infrastructure earns its keep at scale, not below it.
The same retrieval, in production
Quintarthai is our own subscription platform answering Canadian and U.S. equity research questions, where every number cites the filing it came from. The chunking, citation and abstention behaviour described on this page is the behaviour we run there daily — against a corpus that will not sit still.
Our own product, not a client engagement. We have no client case studies yet and will not invent any.
Questions worth asking first
If yours is not here, bring twenty real questions and the documents that should answer them — we test retrieval on a sample before anyone commits to a build.
What happens when the answer isn't in our documents?
The system says so instead of composing a plausible one — an honest "I don't know" when the answer is not in the corpus. That path is also measured: the question set deliberately includes questions whose answer is nowhere, because refusing correctly is a measurable skill.
Can someone get an answer built from a document they may not read?
No. Access control is applied at retrieval, before generation — a user's answer can only be built from documents that user is entitled to read, enforced server-side and never in the prompt. Per-user permission filtering sits in the multi-source tier; the departmental tier covers one corpus and one access tier, and lists per-user filtering as not included.
How do you measure whether it works?
Your subject-matter experts write a question set — typically 80 to 150 real questions — and, for each, the document and passage that actually contains the answer. We then report retrieval and answer quality separately, because one blended score hides which half is broken, and they are fixed by completely different work.
Our documents contradict each other — does that matter?
It is a reason not to buy this yet. Retrieval will faithfully surface both the 2019 policy and the 2024 one that replaced it, and the system has no principled way to know which governs. If your corpus has no versioning and no owner, fixing that comes first. Cleaning up or reconciling contradictory source documents is written down as not included.
Where does it run, and who hosts it?
Hosting on the monthly run tier is in a Canadian or U.S. region, or you run it in your own tenancy. The Canadian regions are AWS ca-central-1 (Montreal) or Azure Canada Central (Toronto); U.S. work runs on AWS or Azure in U.S. regions. Fully on-premise deployment is also available. The region is chosen per engagement and named in the contract before work starts.
Bring twenty real questions
The fastest way to scope this is a handful of questions your team actually asks, and the documents that should answer them. We test retrieval on a sample before anyone commits to a build.
See the assessment