The pipeline, in five steps
Retrieval-Augmented Generation (RAG), with a verification pass on top. Plain English version.
- 01
Retrieve before we generate
Your question is embedded into a vector and matched against the Ghanaian legal corpus we've indexed — the 1992 Constitution, Companies Act 992, GIPC framework (Act 865 + L.I. 1547), Labour Act 651, Rent Act 220, and Supreme Court decisions we've ingested. The model only sees passages that survived this retrieval step.
- 02
Ground the answer in the passages
The model is instructed to answer using only the retrieved passages and to cite the source for every claim. If the retrieved corpus has no relevant material, the model is instructed to say so rather than to confabulate.
- 03
Verify the citations
Once the model produces an answer, each citation it emits is checked against the source documents the retrieval step returned. A citation that points to a passage we can locate is marked verified; one we cannot is marked unverified.
- 04
Show you the count
Every research response carries a count: how many citations were verified, how many weren't. You see it on the response itself and again in the citation panel for the session. Unverified citations are not hidden; you decide whether to trust them.
- 05
Check the citation is still current law
A citation can be real and still be the wrong answer, if the instrument was later amended or repealed. Every cited section is checked against our amendment/repeal graph — hundreds of confirmed relationships between Ghanaian statutes, tracking what replaced what and when. If a citation resolves to a repealed or superseded provision, the response either regenerates around the current instrument or ships with an explicit warning naming what changed and what governs instead. Nothing is presented as current law without this check.
What this protects you from
The single most damaging failure mode of a general-purpose LLM in a legal setting is the fabricated citation — plausible-looking case names that don’t exist, statute sections invented to fit the question. The retrieval step removes the model’s ability to cite anything outside the corpus, and the verification step catches the cases where the model misquotes a real source.
It does not eliminate the lawyer’s review. It narrows the set of failures from “anything the model can imagine” to “the corpus is incomplete or the passage was misread”, and it tells you which one you’re looking at.
Where it can still go wrong
We list the limits up front because the alternative is a user discovering them at the worst possible moment.
- The corpus has gaps. Not every case or regulation is indexed yet. If your question depends on a source we haven’t ingested, the answer will say so — please don’t read the absence of a citation as the absence of a relevant authority.
- Verification is mechanical. A citation counts as verified when the cited passage exists in the source document we retrieved. It doesn’t mean the model characterised the passage correctly. Read the passage.
- Recency lag. The corpus is refreshed on a published cadence; very recent judgments may not yet be present. Treat anything from the last 30 days as uncertain unless you’ve verified it yourself.
- Drafting is a separate pipeline. The drafting workflow takes inputs you provide and produces a structured document. It is not a research surface and does not carry the same citation guarantees.
What you see on every response
- A persistent disclaimer: outputs are AI-generated and must be verified before reliance.
- The citation count, in the form verified / total — for example, 4 / 5 citations verified against corpus.
- The citation panel: each cited source as a clickable entry pointing back to the indexed document.
For the broader trust posture see the Trust hub overview. For the data-handling specifics see the Privacy Notice.