From Zero to Hero: Vector Search Hackday Summary

Group of colleagues in a casual office setting smiling for a team photo — Hackday in Jina AI's Berlin Office, Sept 28th

On Wednesday, Jina AI hosted a Vector Search Hackday with the Haystack EU Community in our Berlin office. Participants were grouped by interests and background, from beginners in vector search to information retrieval professionals.

The "Beginning Vector Search" group built their first Image-to-Text search engine in one day from the ground up, using Jina AI's neural search framework and cloud-based CLIP-as-Service. They were readily able to program a classifier that could distinguish pictures of cats from dogs. After that, the team added a filter for specific dog and cat breeds.

Man in glasses and checkered shirt presenting a cat and dog breeds classifier, with a related code visible on the TV screen — Show & Tell by the "Beginning Vector Search" group

The "Fine-tuning Models" group explored the possibilities of Jina AI's Finetuner. Finetuner makes it easier and faster to optimize pre-trained neural networks for specific tasks. The team first explored the many features of Finetuner, and by the end of the event, they were fine-tuning a network for a specific task.

Group of professionals engaged in a meeting with a woman presenting, laptops and TVs suggesting a tech-focused setting — Show & Tell by the "Fine-tuning models" team

The "Combining TF/IDF & Vector Search" group had a fully running project by the end of the event. For their trial use case, the team used a dataset of 1000 product photos and descriptions. First, they combined images and texts using CLIP-as-Service to get vector embeddings for both. Then, they implemented a comparison of three search algorithms: BM25, CLIP nearest-neighbor, and a hybrid of the first two. Results were promising, and they gained a lot of experience combining text relevance with neural search.

Three individuals engage in a presentation involving a large monitor displaying a webpage, in a room with a whiteboard and a window with greenery — Show&Tell by the "Combining TF/IDF & Vector Search" team

The "Vector Search to E-commerce" group built an end-to-end search engine and deployed it to JCloud. They constructed a Flow with a YAML file and implemented an Executor in it. They were then able to deploy the Flow to JCloud with a one-line command. They were able to set up a monitoring dashboard and by the end of the day had indexed 5000 images with descriptions this way.

Office meeting with individuals focusing on a presentation displaying technical details on a large screen TV — Show&Tell by the "Vector Search to E-commerce" team

To participate in similar events, learn more about Jina AI and meet the team, please follow us on Meetup.

If you want to explore how to build programs like the ones mentioned above, please check out our tutorials and docs.

If you have questions about the Jina AI platform, feel free to join our Slack channel to get answers from our staff.

Join our Slack Community