File size: 4,630 Bytes
bb0e805
 
 
 
 
 
 
 
 
c4301a8
 
4be62d1
 
c4301a8
4be62d1
c4301a8
4be62d1
 
c4301a8
4be62d1
 
 
94b131e
c4301a8
 
 
4be62d1
 
c4301a8
4be62d1
c4301a8
4be62d1
 
16d27ae
 
 
 
 
 
 
c4301a8
 
 
4be62d1
 
 
 
 
94b131e
 
16d27ae
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
---
title: README
emoji: 🌍
colorFrom: green
colorTo: indigo
sdk: static
pinned: false
---

# About Sinequa

Sinequa provides an Enterprise Search solution that lets you search through your company's internal documents. It uses
Neural Search to provide the most relevant content for your search requests.

# Neural Search Models

Sinequa Search uses a technology called Neural Search. Neural Search is a hybrid search solution based on both Keyword
Search and Vector Search. This search workflow implies two types of models for which we deliver various versions here.

The two collections below bring together the recommended model combinations for:

- [English only content](https://huggingface.co/collections/sinequa/best-neural-search-models-for-english-content-673f2d584d396ce427ade232)
- [Multilingual content](https://huggingface.co/collections/sinequa/best-neural-search-models-for-multilingual-content-673f2ec7c6fb004642a24444)

## Vectorizer

Vectorizers are models which produce an embedding vector given a passage or a query. The passage vectors are stored in
our vector index and the query vector is used at query time to look up relevant passages in the index.

Here is an overview of the models we deliver publicly:

| Model                                                                                       | Languages                                       | Relevance | Inference Time | GPU Memory |
|---------------------------------------------------------------------------------------------|-------------------------------------------------|-----------|----------------|------------|
| [vectorizer.vanilla](https://huggingface.co/sinequa/vectorizer.vanilla)                     | en                                              | 0.639     | 14 ms          | 300 MiB    |
| [vectorizer.raspberry](https://huggingface.co/sinequa/vectorizer.raspberry)                 | de, en, es, fr, it, ja, nl, pt, zs              | 0.613     | 12 ms          | 550 MiB    |
| [vectorizer.hazelnut](https://huggingface.co/sinequa/vectorizer.hazelnut)                   | de, en, es, fr, it, ja, nl, pl, pt, zs          | 0.590     | 12 ms          | 550 MiB    |
| [vectorizer.guava](https://huggingface.co/sinequa/vectorizer.guava)                         | de, en, es, fr, it, ja, nl, pl, pt, zh-trad, zs | 0.616     | 12 ms          | 550 MiB    |
| [vectorizer.banana](https://huggingface.co/sinequa/vectorizer.banana)                       | 100+ languages                                  | [details](https://huggingface.co/sinequa/vectorizer.banana#scores) | 35 ms          | 1450 MiB    |

Inference times and GPU memory usage reported are for FP16 models.

## Passage Ranker

Passage Rankers are models which produce a relevance score given a query-passage pair and are used to order search
results coming from Keyword and Vector search.

Here is an overview of the models we deliver publicly:

| Model                                                                                                 | Languages                                               | Relevance | Inference Time | GPU Memory |
|-------------------------------------------------------------------------------------------------------|---------------------------------------------------------|-----------|----------------|------------|
| [passage-ranker.chocolate](https://huggingface.co/sinequa/passage-ranker.chocolate)                   | en                                                      | 0.484     | 13 ms          | 300 MiB    |
| [passage-ranker.strawberry](https://huggingface.co/sinequa/passage-ranker.strawberry)                 | de, en, es, fr, it, ja, nl, pt, zs, zh-trad             | 0.451     | 13 ms          | 550 MiB    |
| [passage-ranker.mango](https://huggingface.co/sinequa/passage-ranker.mango)                           | de, en, es, fr, it, ja, nl, pt, zs, zh-trad             | 0.480     | 65 ms          | 850 MiB    |
| [passage-ranker.pistachio](https://huggingface.co/sinequa/passage-ranker.pistachio)                   | de, en, es, fr, it, ja, nl, pl, pt, zs, zh-trad         | 0.474     | 65 ms          | 850 MiB    |
| [passage-ranker.apricot](https://huggingface.co/sinequa/passage-ranker.apricot)                       | ar, de, en, es, fr, it, ja, kr, nl, pl, pt, zs, zh-trad | 0.449     | 13 ms          | 550 MiB    |
| [passage-ranker.nectarine](https://huggingface.co/sinequa/passage-ranker.nectarine)                   | ar, de, en, es, fr, it, ja, kr, nl, pl, pt, zs, zh-trad | 0.455     | 65 ms          | 850 MiB    |

Inference times and GPU memory usage reported are for FP16 models.