Your documents. Searchable. Cited. POPIA-safe.

RAG Knowledge Base Setup ships the knowledge layer — ingestion, embedding, retrieval, an embeddable chat widget, and the API your applications can call. AI Agent Implementation includes everything in this engagement and a custom agent built on top with tool integrations (Slack, email, CRM, etc.), human-in-the-loop approval gates, and full execution-trace logging. Pick this engagement if your first use case is "chat over my documents". Pick AI Agent Implementation if you need the system to take autonomous actions in other tools.

It is the starting price for the standard scope — up to 4 source connections, up to 5,000 documents, hybrid retrieval, evaluation, embeddable chat widget, 30 days of support. Your final fixed price is confirmed at the end of the scoping phase, before any contract is signed. Once signed, the price does not move unless you formally request additional scope, which is quoted in writing and re-signed before work continues.

Token costs sit on your provider account, not ours, so you have full visibility. Embedding cost is one-off at ingestion plus marginal cost for new documents during sync. Query-time cost depends on volume. As an order-of-magnitude reference: ingestion of a 5,000-document corpus is typically R500–R1,500 in embedding cost; a knowledge base handling a few thousand queries a month against Anthropic Claude typically lands between R1,500 and R5,000 in monthly LLM cost. Both can be capped via budget alerts at scoping.

By default, all production data sits in AWS af-south-1 (Cape Town, South Africa). The Qdrant vector store is self-hosted in af-south-1, document storage is in S3 in af-south-1, and ingestion runs entirely in-region. PII-scrubbed payloads are the only data ever sent to an external LLM provider, and that boundary is enforced by middleware. We can also deliver into Azure South Africa North or GCP Johannesburg if your group security policy dictates a specific provider.

Yes. During scoping we map your user groups to document groups. The retrieval layer enforces this at the chunk level — the LLM never receives a chunk the user is not authorised to see. This is particularly important for legal, HR, and finance use cases where document-level access matters. Access is also part of the audit log: every query records which documents were considered for that specific user.

It depends on your corpus. Recursive character-level chunking is the default for long-form documents (reports, manuals). Sentence-level chunking is preferred for legal text where paragraph boundaries matter. Semantic chunking is used for marketing and editorial content where natural topic shifts dominate. The choice is made during the configure phase based on test query performance — we measure, we do not guess.

Two options. text-embedding-3-small (OpenAI) for the cost-efficient default with strong general performance. A locally-hosted open embedding model (e.g. BGE) when data residency or cost demands it — this runs entirely inside af-south-1 with zero external API calls. The choice is made during scoping with the residency profile of your data in mind.

Priority response on issues, retrieval tuning when answers drift, prompt iteration based on real-world usage, addition of new documents to the index, and Ragas re-runs on request. It does not cover net-new features, additional source systems, or operational on-call — those are quoted as continued support from R8K/month or are included on a platform subscription.

You do. The query API source, infrastructure-as-code (Terraform), CI/CD configuration, the test query set, the chunking and embedding configuration, the runbooks, and the POPIA Section 19 evidence pack are committed to your repositories during the engagement — not at the end. sonofgraig retains no proprietary lock-ins on your knowledge base. If you choose not to convert to a platform subscription, the system runs on standard open-source components your team can maintain.

Yes. sonofgraig has a pre-signed Data Processing Agreement covering processing activities, lawful basis, security controls, sub-processors, and transfer mechanisms. It is available for download from our trust centre and your legal team can mark up departures from the standard text during contracting.

Yes — sonofgraig is B-BBEE certified and CIPC registered. B-BBEE spend certificates are issued per invoice. All commercial documentation is available to your procurement team for supplier on-boarding.