Semantic search finds documents by meaning rather than by matching words, so a search for "staff leaving early" also returns the policy titled "early departure procedure" and the one that never uses either phrase.
It works by turning every passage into a numeric fingerprint — a list of numbers representing its meaning — and doing the same to the query, then returning the passages whose fingerprints are closest. The fingerprints are computed once, when documents are indexed; searching afterwards is fast and cheap. In practice the best results come from combining this with ordinary keyword search, which is still better at exact terms like part numbers, and then re-sorting the combined results with a reranking model.
Models in this group take text or whole documents as input: plain text, PDFs, scanned pages and office files. 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 A4 page of about 500 words indexed, or one query answered against the index.
BGE-M3
Vendor: Beijing Academy of Artificial Intelligence
What it does: turns passages and queries into fingerprints, covering over a hundred languages and passages up to about 8,000 words. It also produces keyword-style scores at the same time, which is what makes a blended search possible from one model.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
8 GB
24 GB
80 GB
vCPUs
6
12
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 (pages/hour)
≈ 6,000
≈ 30,000
≈ 90,000
E5-large / multilingual-E5
Vendor: Microsoft
What it does: a well-proven fingerprinting model, small and quick. A dependable default for English archives, with a multilingual version for mixed collections.
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 (pages/hour)
≈ 9,000
≈ 45,000
≈ 140,000
all-MiniLM-L6
Vendor: Microsoft (Sentence-Transformers)
What it does: a very small fingerprinting model that runs on almost any hardware. Accuracy is below the larger models, but for a modest archive it is fast, cheap and adequate.
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 (pages/hour)
≈ 30,000
≈ 120,000
≈ 400,000
SimCSE
Vendor: Princeton NLP
What it does: a compact model that produces fingerprints suited to sentence-level matching, useful where the unit being searched is a single clause or line rather than a passage.
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 (pages/hour)
≈ 20,000
≈ 80,000
≈ 240,000
BGE Reranker v2-M3
Vendor: Beijing Academy of Artificial Intelligence
What it does: reads a query and a candidate passage together and scores how well the passage answers it. Applied to the top few dozen results, it is usually the single largest improvement in search quality.
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
ColPali
Vendor: Illuin Technology
What it does: fingerprints the page image itself, so slides, scanned forms and charts are retrievable even when text extraction would lose their layout.
Requirement
Minimum
Medium
High
GPU type
RTX 4090
A100 80 GB
2× A100 80 GB
VRAM
20 GB
80 GB
160 GB combined
vCPUs
8
16
32
RAM
32 GB
64 GB
128 GB
Server
1× RTX 4090 24 GB
1× A100 SXM 80 GB
2× A100 SXM 80 GB
Rate (pages/hour)
≈ 900
≈ 3,500
≈ 7,000
Qwen2.5 32B
Vendor: Alibaba Cloud
What it does: turns a vague question into a better query and writes a direct answer over the retrieved passages, which is what turns a list of results into an answer.
Requirement
Minimum
Medium
High
GPU type
RTX 4090 (reduced precision)
L40S 48 GB
2× H100 80 GB
VRAM
22 GB
48 GB
160 GB combined
vCPUs
12
16
48
RAM
48 GB
64 GB
256 GB
Server
1× RTX 4090 24 GB
1× L40S 48 GB
2× H100 SXM 80 GB
Rate (pages/hour)
≈ 350 answers/hour
≈ 1,100 answers/hour
≈ 4,200 answers/hour
Choosing between them
Search quality depends on how your documents are divided into passages, whether keyword search is blended in, and whether a reranker is worth its cost at your volume. Our consultants review your archive and the queries your staff type, then recommend the indexing model, the blend and the reranking step.
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 Semantic search. 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.