Retrieval
Two searches run together — one semantic, one keyword — and their rankings are fused rather than their scores. The two indexes produce numbers on incompatible scales, and averaging them ranks badly. Chunking follows the document’s own structure: Part, Chapter, Schedule, Section, Article, Clause. This is the difference between a citation that reads section 56 and one that reads chunk 47, and the citation is the entire product claim. Repealed provisions are excluded from retrieval rather than filtered out afterwards, so a superseded rate can never be quoted as current. Embeddings are multilingual by choice, so a question asked in Shona can find a passage written in English. That was the deciding factor in choosing the model, and it is now locked: changing it would mean re-embedding the whole corpus and rebuilding the indexes.When Ground does not know
Shamwari says so, answers from general knowledge, and marks that it has done so. It does not invent a section number, a statutory instrument, a rate or a date — and where an answer turns on a current figure it cannot confirm, it says the figure must be verified and names the authority to verify it with. The miss is recorded. A question we could not answer is the most useful thing a customer can tell us about what to ingest next, so the gaps become the corpus roadmap rather than disappearing.Some sources are not ours to take. Every source carries a recorded reuse basis, and the ingestion tool refuses a
source whose terms have not been read rather than warning and continuing.Being able to show something is not permission to train on it. Scraped
news can be cited and linked, and is not training data. Copyrighted exam
material is not scraped at all — licence, or no answer.