Understanding Evidence Ranking Previews

Ranking previews show how StretchSearch would order your indexed candidates as search results. They're produced by a deterministic engine called StretchRank, and every score is broken into transparent components you can audit — no black box.

What a ranking preview is

A ranking preview is a scored, ordered candidate: its type and label, a numeric score, the individual score components, its rank position, a status, and the generation mode that produced it. Previews are read-only artifacts you review — they don't change your site; they show how your content stacks up.

How StretchRank scores candidates

StretchRank (stretchrank_v0.1_deterministic) blends several signals:

• Type weight — entities (1.18) rank above phrases (1.12), above keywords (1.0), above raw chunk references (0.72). More specific candidates count for more.
• Frequency boost — how often the candidate appears, on a logarithmic scale (so repetition helps with diminishing returns).
• Reference boost — how many chunk references back the candidate (capped so a few strong references count, but spam doesn't).
• Evidence boost — how much citable evidence supports it, including evidence confidence.
• Quality penalty — deductions for missing labels or normalized values.

The formula is, in plain terms: base signal × type weight + log(frequency) boost + reference boost + evidence boost − quality penalty. Every preview exposes these components so you can see why something ranked where it did.

Read the ranking overview

1. Open StretchSearch → Ranking previews. The overview summarizes the current previews and their status.
2. Note the persistence state. Ranking previews are currently in a preview/preparation posture — the overview reports that production generation is not yet enabled, so treat previews as a planning and evaluation surface.

Explore the previews list

1. Open the previews list.
2. Filter by candidate type (entity, phrase, keyword, candidate_chunk_ref), ranking status, and generation mode.
3. Sort/scan by rank position. Entities and phrases generally surface near the top.
4. Open a preview to inspect its score components and the source candidate it came from.

How previews are generated

Generation is a governed, operator-only action. When run, StretchRank reads the candidate pool for a project or crawl job, ranks it deterministically, and persists the top results (default 250, up to 500) as active previews in preparation mode. Because it's deterministic, the same candidate pool always produces the same ranking — which is exactly what you want for auditability.

A realistic example

Reviewing Northwind's rankings, the analyst filters to entity candidates and sees Northwind Coffee and single-origin espresso at the top, each with a high type weight and a strong evidence boost. A keyword candidate like coffee ranks lower despite high frequency — its type weight is smaller and it carries a quality penalty. The transparent components make the ordering easy to explain to the client.

Tips

• Read the score components, not just the final number — they tell you whether a candidate ranks on specificity, frequency, or evidence.
• If an important term ranks low, check whether it's classified as a keyword (lower weight) when it should be an entity or phrase.
• Strengthen a candidate's evidence boost by ensuring it's backed by citable evidence references.
• Because generation is deterministic, re-running on the same pool won't change results — regenerate after the underlying candidates change.

FAQ

Can I edit a ranking preview's score? No. Previews are computed artifacts. Change the inputs (candidates, evidence) and regenerate.

Why is production generation shown as not enabled? Ranking-preview persistence is in a governed preview stage. Use previews for evaluation and planning; your platform operators control when generation is enabled.

Is this the same ranking used in federated search? No. Federated search uses a separate lexical/freshness/authority ranker across result types. StretchRank scores candidates for retrieval and evidence purposes.