Use case
Finding sources worth following
Teams building recommendation, curation and discovery products.
The problem, and the approach
The problem
- Discovery by keyword returns what matches the words, which is not the same as what is worth reading.
- Widening a result set usually means guessing at new queries, which reflects your vocabulary rather than the corpus's.
- Popularity signals favour recency and volume over substance.
The approach
- sort=centrality orders by position in the coverage graph — what the field references, rather than what matched the query.
- similar widens from a seed passage through the embedding space, crossing languages as it goes.
- related follows typed edges, so the expansion comes from documented relationships rather than similarity alone.
- Community filtering keeps discovery inside one discourse when a term means different things in different fields.
The parameters that matter
| Parameter | Value | Why |
|---|---|---|
sort | centrality | Foundational rather than merely matching. |
similar?id= | a good seed | Widen without inventing a query. |
min_centrality | 0.3 | A floor on how well-connected results must be. |
community[] | one cluster | Stay inside a single discourse. |
curl -H "x-api-key: $UNLOB_API_KEY" \
"https://api.unlob.com/search?q=retrieval+evaluation&sort=centrality&limit=10"
curl -H "x-api-key: $UNLOB_API_KEY" "https://api.unlob.com/similar?id=p:8f2c9a&limit=10"Centrality is a better editor than recency
Discovery products that rank by recency converge on churn, because the newest thing in any field is almost never the most useful thing in it. Ranking by how load-bearing a passage is within its field — how much of the surrounding material depends on it — produces a different and generally better shortlist, and it is computed from the structure of the corpus rather than from engagement, so it does not degrade the way behavioural signals do when the audience is small.
The cost is a conservative bias. Centrality accrues over time, so genuinely new material is under-served by construction, and a discovery feed built on it alone will feel authoritative and slightly stale. Pairing it with a recency preference rather than a recency filter keeps the established work available while letting new work surface.
Widening from a seed beats inventing queries
The hardest part of a recommendation surface is the second step: the user liked this, now what. Generating new queries from the text of the seed is the obvious approach and it produces the characteristic failure of recommendation systems, which is returning things that are lexically similar and substantively the same.
Expanding along relationships the corpus already contains avoids that, because the edges were derived from how the material actually connects rather than from how it is worded. Two practical notes: community assignment is automatic and occasionally splits a field a person would treat as one, which shows up as a discovery feed that is oddly narrow; and a seed from a very small community will widen into a thin set, so it is worth probing the size of the neighbourhood before presenting the result as a recommendation.
What a fair test looks like
A representative task: Recommend related material for the item a reader is viewing. Take twenty of yours, write down what a good result is before running anything, and score them on:
- Recommendations are related to the item, not near-duplicates of it.
- Sources span more than one publisher or community.
- Recommendations respect your freshness and content-type filters.
Include tasks where the right answer is to abstain, and report the cases where unlob does not fit alongside the ones where it does. How an evaluation works.
What it costs
The recipe above, as an illustrative call pattern: 2 similar + 1 related + 1 searchper recommendation set — 5 credits. Two similarity walks from the item in view, one related lookup and one search to fill the set.
| Tasks a month | Credits | Cheapest plan | Per 1,000 tasks |
|---|---|---|---|
| 1,000 | 5,000 | Build, $29/mo | $29 |
| 10,000 | 50,000 | Build, $29/mo | $2.90 |
| 100,000 | 500,000 | Scale, $199/mo | $1.99 |
Provider spend only; your model and review effort are separate lines. The pattern is a starting point, not a measurement — cost your own.
Where this is not the right tool
- Centrality favours established sources, so it under-surfaces genuinely new material. Combine with prefer_recent when novelty matters.
- Community assignment is automatic and occasionally splits a field that a person would treat as one.
Frequently asked questions
How is centrality different from relevance?
Relevance asks what matches your query. Centrality asks what the corpus treats as load-bearing, independent of the query. For discovery the second is usually the better ordering.
Does similar work across languages?
Yes. The embedding space holds 101 languages, so a German seed passage can surface its English counterparts directly.
Evaluate it on twenty of your tasks
Send the tasks and what a good result looks like, and we report what was found, what was missing, the cost and where unlob does not fit — or run it yourself on 10,000 free credits a month.