<aside> ✨

Online Art Gallery

Key Metrics

One million unique artworks in catalog

Agent implemented

Chatbot driving customer down the conversation funnel, from discovery to repeat purchases

</aside>

Context

Singulart enlisted us for a special experiment: creating an agent that can guide visitors, depending on their context (newcomers, repeat buyers…) and persona (based on their browsing behaviour and history with Singulart), through Singulart’s catalog of one million plus unique artworks.

Achievements

🔬 Phase 1: Scoping

Specifications, Data Sources, Data Lifecycle, Integration

The main technical challenge was to implement a RAG (Retrieval Augmented Generation) approach that would perform for Singulart. For an efficient RAG implementation, there are three parts:

• preparing the data (data lifecycle, chunking, vectorizing)
• creating the search function
• re-ranking the results, as well as other techniques.

There concepts are thoroughly explained by our co-founder Jonathan in a talk he gave in October 2023. A video is worth 24 images per second, each worth a thousand words… (the talk was in French, the video is in French, but you can access subtitles):

https://www.youtube.com/watch?v=tnpZYr4isBE

🛠️ Phase 2: Implementation

Focus: Prompt Engineering, Functions, Automations, User Experience, Testing

When implementing the assistant, we made numerous iterations on the user experience for the agent:

• it should not replace the perfectly tuned navigation experience of the website, but be part of it
• artworks should be highlighted in the context of the conversation by the assistant, while leveraging the mastery of showcasing the artworks of the website.

👁️ Phase 3: Optimisation

Focus: Usage, Iteration, Support

For the beta-testing phase, the assistant has been deployed to the largest English-speaking Singulart audience, the American market.

We carefully monitored the performance of the agent across metrics such as:

• sentiment analysis: did the conversations feel right?
• performance of the search
• usage metrics
• conversion metrics, as in any ecommerce case.

Conclusions

🎯 Results & Learnings

• Semantic search, achieved through the RAG technique, performed perfectly well, as 98% of searches human-reviewed provided artworks related with the search
• Most interesting were the learnings about usage: visitors who engaged in a conversation with the agent kept interacting and visited back… but only 5% of visitors started an interaction, and 50% of these interacted with using single words for their search instead of writing sentences, which defeats the purpose of semantic search and conversational agents.

🛤️ Aftermath

• These results suggested that there is a long road before the general public starts forgetting about first generation chatbots, in which decision trees often led to a wall. There are two options to move further with the experiment: providing the agent with contextual proactivity; or create a dedicated (mobile) app, designed with LLMs at its core.
• Singulart is now experimenting with agents to automate internal processes.