TotalApp Docs
Ctrl K
EN | TR
Go to App

Reranking

A second look at the shortlist — a dedicated model re-scores the best embedding matches so the most relevant record rises to the top.

Overview

Embedding is fast but approximate: it compares meanings from a distance. Reranking is slower and sharper. A rerank model reads your text together with each shortlisted record and scores how well they truly fit, then reorders the list. Near-misses stop outranking the real answer.

Optional, and never in the way

Reranking is an optional refinement. If no rerank model is selected, or the service is unavailable, you simply see the normal embedding order — no error, no broken screen.

How it works

Embedding shortlist
→
Best candidates sent to the rerank model
→
Each re-scored against your text
→
List reordered
→
Top matches kept
  1. Embedding produces a shortlist of candidates above the screen’s minimum similarity.
  2. Up to the 30 best candidates are sent to the rerank model together with your text.
  3. The model returns a relevance score for each one.
  4. The list is reordered by that score and only the strongest few are kept.
EmbeddingReranking
Question it answers“Which records are roughly about this?”“Which of these fits best?”
How it readsText and records separatelyText and each record together
SpeedFast across many recordsSlower, so it only sees the shortlist
Runs onLocal Ollama or a hosted modelA hosted model
RequiredYes, for semantic rankingNo — optional

Turn it on

  1. Make sure an embedding model is selected in Settings → Agentic → Embedding.
  2. Open the Rerank sub-tab. Tick OpenRouter: free models only if you want a no-cost model.
  3. Click a model — for example NVIDIA: Llama Nemotron Rerank VL 1B V2 (free).
  4. To turn Reranking off again, click the selected model once more.

What you will see

  • A green Reranked label above the results when the second pass was applied.
  • The percentage on each row is then the rerank score, not the similarity score — so percentages from the two orderings should not be compared with each other.
  • Reranking only runs when at least two matches clear the minimum similarity. With one match, or none, there is nothing to reorder and it is skipped — that is normal, not an error.

Models

Rerank models are read live from your providers, so new ones appear without an app update. Examples available today:

ProviderExample modelsGood to know
OpenRouterNVIDIA Llama Nemotron Rerank VL 1B V2 (free)Free to try; about 1 billion parameters, fast
OpenRouterCohere Rerank 4 Pro, 4 Fast and v3.5; VoyageAI rerank-3 and rerank-2.5; Qwen3 Reranker 8BPaid options with higher capacity
Modelionqwen3-reranker-0.6bGateway model, no separate key to manage

Some rerank models can also read images. TotalApp currently sends text only.

Privacy

Hosted only

Unlike a local Ollama embedding model, Reranking sends your query and the shortlisted records’ text through the TotalApp server to the provider of the model you picked. If that text must not leave your environment, use a local embedding model and leave Reranking unselected.

Where it is used today

Reranking is available on Trade Compliance & Sanctions: after Search Similar Rules shortlists the sanction rules that fit a shipment, the rerank model reorders them before you choose which rules go into the AI diagnostic. It will be extended to more screens over time.

Frequently Asked Questions

Does Reranking need Embedding?
Yes. It re-scores the shortlist that Embedding produces, so an embedding model must be selected first.
Why did I not get a Reranked label?
Usually fewer than two matches cleared the minimum similarity, no rerank model is selected, or the rerank service was unavailable or over its quota. In every case the normal embedding order is shown.
Is it free?
Free rerank models are available. Paid models are billed by the provider; free models may have a daily request quota.
Will it change the AI diagnostic?
Only indirectly: a better-ordered shortlist means the records you keep selected are more likely to be the right ones, so the AI step reasons over a sharper context.