bgevconsultancy is making Helm-v2's document search far more exact
This fork treats retrieval quality as a product capability, not a background technical detail.
bgevconsultancy has rebuilt the machinery that finds the right material before an AI answer is formed, with particular attention to keeping legal text meaningful in context.
- Self-hosted document representations reduce reliance on an outside service, though they require the team's own infrastructure.
- Clause-aware document splitting aims to keep provisions and sections together rather than cutting through them.
- Organisation-scoped search keeps retrieval separated by organisation.
- A second-pass relevance check reorders initial results to surface the strongest matches.
- A repeatable evaluation set measures search quality instead of relying on informal spot checks.
The fork also accounts for the delay while a self-hosted model starts up, a small but useful sign that the team is thinking about the working experience as well as accuracy.
So what Legal teams building contract or knowledge search should care because better retrieval is often the difference between a persuasive answer and a confidently wrong one.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?