How to choose search for a support knowledge base.

You work in a connected AI client, such as Codex, ChatGPT or Claude. Dovod holds the project’s research materials, sources and findings so you can review the work and continue later.

For this example, bring a description of your knowledge base, a few unsuccessful searches and your technical constraints. The first outcome is a clear evaluation plan; a final technology choice needs evidence from your own trial.

It would be easy to assume typo handling requires a new engine. PostgreSQL’s documentation describes using pg_trgm with full-text search for misspelled input. That gives us a baseline to test, rather than a reason to replace it immediately.

Meilisearch documents typo tolerance and indexing tasks. Typesense documents query-level typo controls. None of those pages tells us which engine ranks our articles best.

Check the PostgreSQL source ↗

Before any candidate was read, a clarification session fixed the gates: German is a hard requirement, access restrictions must hold, and nobody names a freshness number the team has not agreed. After the findings, a review session asked whether the documentation justified a choice. It did not — and that is recorded, answer by answer, with links to the material each answer rests on.

One question is deferred and one is still pending. A session that ends with everything answered would be a script; this one ends where the evidence does.

The matrix puts the three candidates against the agreed criteria and refuses to score them. Each cell says one of three things: the vendor's documentation describes the mechanism, we inferred it, or nobody has tested it. Most of the relevance and cost rows are “not tested” — which is the honest state of the project, and the reason the decision record asks for a trial.

Criteria, sources and review are done. The trial is not, and the decision record cannot be revisited until it runs. Each step on the roadmap points at the document, session or task it stands for, so the plan and the evidence are the same project — not a slide and a folder.

The same project also draws itself as a knowledge graph and a mind map: which findings the ADR cites, which questions a finding answered, which tasks are waiting on which document. Open the live project and choose Views.