Loading...
Loading...
Services / Knowledge Systems
Staff and clients ask questions in plain language. The system answers only from your contracts, policies, and operations docs — with the exact section cited every time. Never hallucinated. Never exposed to the public web. Auditable: every answer is verifiable against the source.
Book a CallWhat it actually does
A Knowledge System turns a pile of documents into a searchable, conversational knowledge base. Staff or clients ask questions in plain language. The system answers only from your documents and cites the source for every answer. The architecture is retrieval-based: your documents are indexed, retrieved by similarity at query time, and the language model is constrained to answer only from what it retrieved. If the answer isn’t in your documents, the system says so.
Document quality is the input that determines output quality. Whether contracts have clear clause numbering, whether policy documents have proper headings, whether technical manuals separate procedure from explanation — the structure of your source content shapes how cleanly the system can retrieve the right section for the right question. We assess this upfront in the Map stage so there are no surprises in Production.
The stack we build on is conservative and well-understood. Open-standard vector storage, established retrieval, the leading reasoning model for your domain and language. Hosting defaults to SA or EU regions for POPIA compliance. You own the database, the code, and the credentials at handover.
When to use this
Not a fit if
What’s included
How it’s delivered
Each stage is fixed-price and self-contained. The Map stage is free. The Pilot proves search quality on a real subset of your corpus before the full ingest. If the Pilot doesn’t earn its place, we don’t proceed to Production.
Confirm the corpus shape — volume, formats, where the documents live — and run a quality check on a sample of your files. Identify the audience (staff, clients, or both) and the right interface. Set the language scope: English only or bilingual Afrikaans/English. Decide on update cadence: one-time ingest of a static corpus, or a live sync with a moving document store.
Working search over a representative subset of your corpus. Your team runs real questions and verifies the answers are accurate and well-cited. Pilot deliverable: a working interface, an accuracy report against real test queries, and the full Production scope. If the Pilot doesn’t earn its place, we don’t move on.
Process the entire document corpus. Search-quality hardening against a real-question test set drawn from your team. Production interface build with authentication and role-based access control. Citation rendering throughout. POPIA review, training, handover.
Knowledge Systems drift quietly. New documents arrive; old ones get superseded; phrasing of questions shifts. The Run engagement covers incremental ingest of new documents, periodic accuracy review against a refreshed question set, and model migrations when better options ship.
Why this approach
The case for citation discipline isn’t aesthetic. Language models that aren’t forced to cite will confabulate confidently — and confident wrong answers are worse than “I don’t know” in a regulated practice.
visibility lift for pages that add named citations to credible sources. The same principle applies inside a Knowledge System — cited retrieval beats uncited. Aggarwal et al., Princeton, KDD 2024.
of generative-AI projects forecast to be abandoned after proof-of-concept — AAA’s Pilot gate exists to catch this before the full corpus spend. Gartner, July 2024.
accuracy bar agreed in writing before Production begins. The threshold is set against questions drawn from real staff or client queries on your corpus — not synthetic tests.
A well-built agent will handle most inputs correctly. The review queue handles the rest. If you need 100% accuracy, you need a human — not an AI.
Matt Owen, in From ChatGPT to Custom AI Agents.
The same logic governs Knowledge Systems. Citation rendering is the architectural fix for the “I don’t know what I don’t know” problem — the user can always check the source. AAA’s defaults map to the standalone-extractable principle from Aleyda Solis: every answer should stand on its own, with the source hooks a reader needs to verify it.
Common questions
ChatGPT answers from its training data — generic, sometimes wrong, and never includes your contracts or policies. A Knowledge System answers only from your own content, cites the exact source for every answer, and refuses to confabulate when the information isn’t there. It is auditable: you can verify every answer against the document and section it cited. None of your content is exposed to other ChatGPT users or used for model training.
Accuracy depends on three things: document quality (clean text beats scanned images every time), how your source content is structured (well-organised documents retrieve better than freeform notes), and the corpus itself (if it’s not in your documents, the system says so rather than guessing). The Pilot stage proves accuracy against your team’s real questions before Production commits.
Yes — the leading reasoning models handle Afrikaans business text well, and modern retrieval works across both languages. For corpora that mix English and Afrikaans (Western Cape wine estates, Boland agricultural businesses, Paarl property firms), we test thoroughly in Discovery and account for bilingual prompt tuning in the Pilot.
Every engagement includes a POPIA review with a written data-flow diagram. The default hosting choice is SA or EU regions — both acceptable for POPIA Section 72 cross-border transfer. LLM providers operate under their published Data Processing Agreements for business customers, and PII can be stripped or pseudonymised at ingest where the document type allows it. For special-category data (health records, attorney-client privileged content), we scope strictly and document the boundaries explicitly in the SOW.
Tell me about the documents your team has and the questions they keep getting asked. I’ll tell you whether a Knowledge System is the right shape — or if a simpler tool will do.
Book a Call