RAG development for Canadian teams that need cited answers.

Eastern Canadian hours (Toronto, Montreal, Ottawa) get a native two-to-three-hour late-afternoon overlap with our Mohali IST day, which is workable but not full coverage. For Eastern clients that need full Bay Street business-hours coverage, we route a dedicated overlap pod through our Frisco, TX office — Frisco runs Central Time, one hour behind Toronto but covering the same workday. Western Canadian hours (Vancouver, Calgary) are thinner; we cover them async-first with twice-weekly demos in Pacific morning showing retrieval-recall and citation-faithfulness numbers. Written async updates with eval-run numbers go out daily before your standup, so you walk into the day already knowing what regressed overnight.

Yes. Engagement defaults align with PIPEDA federally and Quebec Law 25 for any RAG processing Quebec-resident personal information. Every retrieval and generation call is audit-logged with query, retrieved passage IDs, citation faithfulness score, prompt version, and operator identity — exportable for a Privacy Commissioner inquiry or a Commission d'acces a l'information review. A Privacy Impact Assessment (PIA) is run for any RAG processing personal information at scale or operating on sensitive categories. The refusal layer is wired in week one with a measurable out-of-scope rate so the system never fabricates an answer when the corpus is silent. For Article 12.1 of Law 25 (automated decision-making transparency), the citation and confidence score give the data subject the information they are entitled to.

Your call. We default to AWS ca-central-1 (Montreal), Azure Canada Central (Toronto), or GCP northamerica-northeast1 / northeast2 for Canadian clients, and we will run the entire build inside your Canadian cloud account if your DPO requires no cross-region replication and no data egress to US endpoints. The vector index (pgvector, Qdrant, or Weaviate) lives where you specify. For LLM inference, we pin Claude or GPT-4o to a Canadian or US region depending on what your DPA permits, or we self-host Llama 3 on vLLM inside your VPC for zero third-party inference. Embedding models can be Canadian-hosted for clients who refuse to send corpus passages to a US endpoint.

A PIPEDA + Law 25-aligned Data Processing Agreement covering processor obligations: processing only on documented instructions, confidentiality of personnel, security of processing, sub-processor management, breach notification, and deletion or return of personal information at the end of the engagement. Mutual NDAs are signed before any technical detail or sample corpus is shared. For healthcare RAG (PHIPA in Ontario, HIA in Alberta), provincial-equivalent processor agreements are signed before any PHI is shared. For OSFI-supervised clients, our DPA includes the documentation required under Guideline B-10 (sound business and financial practices) for third-party arrangements. Schedule 4 spells out the safeguards in place for any cross-border processing.

Either. Most repeat Canadian clients move to a Master Services Agreement after the first engagement so subsequent RAG builds, evaluation work, and on-call retainers ship under a per-project Statement of Work without renegotiating the umbrella terms. For a first engagement, a standalone SOW with the DPA appended is the standard pattern. Legal turnaround is usually one to two weeks depending on your DPO and procurement review cadence; we work from your legal team's MSA template or provide ours.

Because pure vector retrieval drops obvious keyword matches that legal, financial, and clinical users in Canada notice immediately. The classic failure is a user searching for an exact statute citation, a fund code, a CUSIP, a billing code, or a specific drug name — and the vector model returns a semantically similar but lexically wrong document. Hybrid retrieval (BM25 for high-precision keyword matches plus dense vectors for semantic recall, blended via reciprocal rank fusion) gives both. It is the default we ship for Canadian legal, financial, and healthcare RAG because regulated users will not accept a system that misses the literal phrase they searched for.

Most v1 RAG engagements at Aiinfox land between CAD $40,000 and CAD $160,000 fixed-price for a focused build — a financial RAG, a medical-inquiry RAG, a legal research copilot, or a knowledge-base copilot. Larger multi-quarter engagements with bespoke embeddings, custom evals, Law 25 documentation, and integration into a regulated platform typically reach CAD $200,000 to CAD $360,000. The cost difference versus a Bay Street consultancy lands roughly 30 to 50 percent lower on senior rates — useful, but the headline is the engineer on your kickoff call writes your retrieval pipeline through launch, with no swap-out to a junior pool mid-engagement.

Yes — takeover audits are routine. Step one is reading the ingestion code, the chunking strategy, the retrieval evaluation results (if any exist), the prompts, and the cost telemetry. Step two is shipping the smallest valuable change — usually a hybrid-retrieval upgrade or a proper citation-faithfulness eval — to prove we understand the system. Step three is the longer-term plan: incremental stabilisation, a parallel rebuild on hybrid retrieval, or shutting it down and starting over. Most takeovers we see did not need a full rewrite; they needed evals, hybrid retrieval, a refusal layer, and a senior engineer on the build.