How we measure
The numbers we have, and the ones we do not.
Every figure here is measured. Where a figure does not exist yet, this page says so and leaves the space empty, because a plausible number is worse than no number.
- The corpus. Built in Baltics staging: 525 members, 24 of them with any evidence at all. A corpus where everyone looks great is one you would be right to disbelieve.
- Response time. Under three seconds, currently met. Above that it stops being the product and becomes a report you wait for.
What pure semantic search does on that corpus
Embed every profile, embed the question, return the nearest. Same corpus. It returns claimants over practitioners by construction, and everything Keegi does differently was built after watching this happen on real data.
Asked
who has machine learning experience?
The best-evidenced member ranked 45th. The top 20 all had an evidence density of zero, so the system answered “nobody”.
Asked
who knows building saas products
Three genuinely qualified people existed. One reached the ranker, at rank 16 of 20. The 15 above it were near-identical “bootstrapping a B2B SaaS” bios. The right answer arrived by luck.
The two questions we cannot answer yet
There are exactly two things worth knowing about whether this works. First: when the right person exists, do we find them at all? Second: when we find them, do we put them in the three names you actually see? The first failing means retrieval is broken; the second means the ranking is.
We measure both internally on every change. They are the numbers behind any claim we could make about quality. We have not published them, because we have not run them on a corpus we would name in public. When we do, they go here. Until then, treat this page as incomplete, because it is. If you want to ask about them by name in a technical review, they are recall@20 and hit@3.
The models
Claude Sonnet writes the capability cards and does the ranking. text-embedding-3-small does retrieval. Model ids are one line of config behind a gateway, so this list can change. When it does, it changes here.
What Keegi is bad at
- A thin corpus. If your people have not written, shipped or resolved things somewhere Keegi can read, there is nothing to ground a sentence in, and Keegi will say so rather than guess.
- Brand-new hires. Someone three weeks in has produced almost nothing internally, so Keegi will not find them, no matter how good they are.
- Capabilities that only ever existed in someone’s head. If it was never written down, Keegi cannot see it. Neither can anything else.
- Emphasis. Keegi will occasionally be right that someone did a thing and wrong about how central it was to them. That is why the answer is three names and a human decision, not one name and a routing rule.
Twenty minutes. No demo, no pitch. I want to know whether this problem is real where you work.
Five questions firstBack to the demo