How your agent finds answers
Every answer your agent gives is retrieved from content you own. It doesn't answer from general world knowledge, and it doesn't make things up when it finds nothing.
Two kinds of search, fused
Most search is either keyword-based or semantic. Each fails in a different way, so HelpShelf runs both and merges the results.
Keyword search matches literal words. It's excellent for exact terms — error codes, field names, product SKUs — and useless when the visitor uses different words than you did.
Semantic search matches meaning, using a numeric representation of each article called an embedding. A visitor asking "why was my card rejected" finds your article called "Failed payments" even though they share no words. It's weaker on exact strings.
The two ranked lists are merged with Reciprocal Rank Fusion: an article that both methods rank highly ends up top, and an article that only one method found can still surface. You get the precision of keyword search and the flexibility of semantic search at once.
Trust tiers change the ranking
Not all content deserves equal weight. Every item carries a trust tier that multiplies its score:
| Tier | Weight | Where it comes from |
|---|---|---|
| Curated | 1.3× | Articles you wrote by hand, or imported from your help desk |
| Standard | 1.0× | Pages picked up by the website scan |
| Generated | 0.8× | Drafts your agent wrote itself |
So a hand-written article beats a scraped marketing page on the same topic, and an AI-drafted article ranks below both. See Trust tiers.
Confidence and citations
Each answer carries a confidence level derived from how strong the retrieved matches were. Answers cite their sources as numbered references — click one to read the original article.
If confidence is low and the visitor has already asked a few times, the widget offers to hand them to a human instead of trying again. See When your agent escalates.
Why an article might not be found
In rough order of likelihood:
- It has no embedding yet. It'll be findable by keyword but not by meaning. Check AI search coverage on your readiness score.
- It's unpublished. Unpublished content is excluded from retrieval.
- The wording is too far off. If nothing in the article resembles how customers ask, add a line that uses their words.
- Something else outranks it. A curated article on a near-identical topic can crowd it out.
Improving answer quality covers what to do about each.