Spaces:
Running
Running
| 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 | 53 ms | 330 MiB | | |
| | [vectorizer.raspberry](https://huggingface.co/sinequa/vectorizer.raspberry) | de, en, es, fr, it, ja, nl, pt, zs | 0.613 | 52 ms | 610 MiB | | |
| | [vectorizer.hazelnut](https://huggingface.co/sinequa/vectorizer.hazelnut) | de, en, es, fr, it, ja, nl, pl, pt, zs | 0.590 | 52 ms | 610 MiB | | |
| | [vectorizer.guava](https://huggingface.co/sinequa/vectorizer.guava) | de, en, es, fr, it, ja, nl, pl, pt, zh-trad, zs | 0.616 | 52 ms | 610 MiB | | |
| | [vectorizer.banana](https://huggingface.co/sinequa/vectorizer.banana) | 100+ languages | [details](https://huggingface.co/sinequa/vectorizer.banana#scores) | 35ms (fp16) | 1450 MiB | | |
| ## 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 | 64 ms | 550 MiB | | |
| | [passage-ranker.strawberry](https://huggingface.co/sinequa/passage-ranker.strawberry) | de, en, es, fr, it, ja, nl, pt, zs | 0.451 | 63 ms | 1060 MiB | | |
| | [passage-ranker.mango](https://huggingface.co/sinequa/passage-ranker.mango) | de, en, es, fr, it, ja, nl, pt, zs, zh-trad | 0.480 | 358 ms | 1070 MiB | | |
| | [passage-ranker.pistachio](https://huggingface.co/sinequa/passage-ranker.pistachio) | de, en, es, fr, it, ja, nl, pl, pt, zs, zh-trad | 0.380 | 358 ms | 1070 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 | 64 ms | 1100 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 | 369 ms | 1200 MiB | | |