Open-source AI search plugin Evoke: 30 million parameters, recall rate close to 600 million models.

Beating AI News Flash: Stability AI co-founder Emad Mostaque's new company Intelligent Internet has open-sourced Evoke. This is a PostgreSQL database search plugin with a built-in AI model of approximately 30.3 million parameters. After installation, developers can search documents and knowledge bases directly within the database. It can match keywords as well as find content with similar meanings.

Evoke is based on IBM's Granite-Embedding-30M-Sparse model. It generates a set of weighted related terms for documents, making it possible to find corresponding content even if the search terms do not appear in the original text. Keywords and semantic information share a single search index, with no need to maintain an additional vector index. The model runs directly on an ordinary CPU, requires no GPU, and does not need to send data to external AI services.

Previously published official English retrieval tests showed that Evoke achieved a relevant document recall rate of 66.7% within the top 100 results, higher than traditional keyword search's 56.3% and close to the 67.1% of a 600 million-parameter model.

Evoke's code and model are both open-sourced under the Apache-2.0 license, making it suitable for scenarios such as enterprise internal document retrieval and AI Agent knowledge bases.