Reranking takes the results your search already returned and puts them in a better order, by reading each candidate together with the query instead of comparing precomputed fingerprints.
The distinction matters. An index must be able to compare a query against millions of passages, so it compares summaries of them — fast, but approximate. A reranker looks at only the top twenty or fifty candidates and can afford to read each one properly against the query, which catches the cases where a passage looked relevant but is not. It is the cheapest large improvement available to an existing search system, because nothing else has to change: the index stays as it is and the reranker is inserted in front of the results.
Models in this group take a query and a candidate passage together as input, and return a relevance score. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one query scored against one candidate passage of about 200 words.
BGE Reranker v2-M3
Vendor: Beijing Academy of Artificial Intelligence
What it does: reads query and passage together and scores the match, across a hundred languages and long passages. The strongest general-purpose reranker we deploy, and the usual first choice.
Requirement
Minimum
Medium
High
GPU type
RTX 3090
RTX 4090
A100 80 GB
VRAM
12 GB
24 GB
80 GB
vCPUs
8
12
24
RAM
32 GB
48 GB
96 GB
Server
1× RTX 3090 24 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (query-passage pairs/hour)
≈ 40,000
≈ 150,000
≈ 400,000
DeBERTa v3
Vendor: Microsoft
What it does: a small reranker fitted to your own click and feedback data, so it learns which results your staff actually found useful rather than a general notion of relevance.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
6 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (query-passage pairs/hour)
≈ 60,000
≈ 240,000
≈ 700,000
all-MiniLM-L6
Vendor: Microsoft (Sentence-Transformers)
What it does: a very small reranker for cases where added latency must stay under a few milliseconds. Less accurate than the larger models but almost free to run.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
4 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
8 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (query-passage pairs/hour)
≈ 200,000
≈ 800,000
≈ 2,400,000
E5-large / multilingual-E5
Vendor: Microsoft
What it does: used in a lighter reranking mode, comparing query and passage fingerprints computed at query time. A middle option when a full reranker is too slow.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
6 GB
24 GB
80 GB
vCPUs
4
8
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (query-passage pairs/hour)
≈ 90,000
≈ 400,000
≈ 1,200,000
Mistral Small 3
Vendor: Mistral AI
What it does: reads the whole candidate set at once and orders it, which handles queries where the best result only becomes clear in comparison with the others. Considerably slower, so it is reserved for small candidate sets.
Requirement
Minimum
Medium
High
GPU type
RTX 4090
L40S 48 GB
H100 80 GB
VRAM
24 GB
48 GB
80 GB
vCPUs
12
16
32
RAM
48 GB
64 GB
128 GB
Server
1× RTX 4090 24 GB
1× L40S 48 GB
1× H100 SXM 80 GB
Rate (query-passage pairs/hour)
≈ 4,000
≈ 12,000
≈ 36,000
Choosing between them
The trade-off is quality against added latency, and it depends on how many candidates you rerank and how quickly results must appear. Our consultants measure the gain on your own queries, then recommend a model and a candidate count that fit your response-time budget.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for Search reranking. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.