Data flow
When a user describes a segment to build, our system gathers the relevant catalog from internal data stores and aggregates it into a (JSON) LLM-readable format. After precise prompt tuning and elaborately crafted flows, we provide the AI with the user’s perceived intent, reference information, and plenty of sanity checks. Once the semantic intent is understood, the AI queries the metadata from the backing databases and identifies the relevant entities. Each of these entities is refined via a business context aware, exhaustive set of sanity checks through custom tailored heuristics to keep the agent’s hallucinations in check. Special care is taken to ensure that these processes do not change or interfere with the underlying client data.
Vector similarity search
Vertex AI’s Vector Search employs machine learning to grasp the essence and context of disorganized data. It relies on huge pre-trained models (text-embedding-gecko in this case) that have a broad range of knowledge and interpret meanings with great accuracy. These models can translate words, sentences, or paragraphs into numerical representations. These numerical representations encapsulate the root meaning, and as a result, similar numbers match similar ideas.
To break down the task into operators and values, we make two distinct requests to the Vertex AI’s LLM PaLM2 (text-bison) using the same basic information. Each request involves the context (user’s input), available values, and the previous agent’s response within the input. Because the pool of operators is limited, the AI Companion can consistently provide a reasonable response without needing further refinement. However, the actual answer may be wrong or missing. Additionally, the agent can only use a limited range of values and doesn’t understand columns with multiple possibilities. To address this, we compare Ada’s answer to the possible value for that column until we find a match using similarity search.
Once we have built these three structured groups, the primary role of PaLM-2 text-bison concludes. At this point, we employ these well-structured groups to construct SQL queries, which run using a designated client SQL Query Engine. We use this output to pre-fill the segment conditions which the user can verify and save the audience for activation.
Better together: Zeotap + Google Cloud
Zeotap and Google Cloud are working together to transform how companies manage and engage with their clients. Our collaborative solutions are already driving value for Zeotap’s customers, offering an innovative, user-friendly interface that prioritizes simplicity and results. By harnessing the power of Google Cloud’s gen AI models, we are committed to making data-driven marketing more accessible and efficient.
Google’s generative AI technology has been instrumental in helping us unlock new possibilities for our customers. The synergy between Zeotap’s platform and Google’s advanced models has enabled us to deliver innovative solutions that improve accuracy, efficiency, and personalization. We are excited to continue collaborating with Google and exploring the potential of generative AI to transform the industry.
Our vision extends beyond audience refinement; we are dedicated to enhancing user experiences and pioneering innovative solutions. This involves streamlining data integration, automating data mapping, and equipping marketers with cutting-edge AI technology for effortless customer insights. In the upcoming months and years, Zeotap is committed to continuing our collaboration with Google Cloud to capitalize on all of the benefits that gen AI will bring to our customers.
Learn more about Google Cloud’s open and innovative generative AI partner ecosystem. Read more about Zeotap and Google Cloud.
References
- Chen, Zixiang, et al. “Towards understanding mixture of experts in deep learning.” arXiv preprint arXiv:2208.02813 (2022).
- Karpas, Ehud, et al. “MRKL Systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning.” arXiv preprint arXiv:2205.00445 (2022).
- Yao, Shunyu, et al. “React: Synergizing reasoning and acting in language models.” arXiv preprint arXiv:2210.03629 (2022).
- Ji, Bin. “VicunaNER: Zero/Few-shot Named Entity Recognition using Vicuna.” arXiv preprint arXiv:2305.03253 (2023).
