How Conversational agent uses AI

This guide covers how Conversational agent uses AI while maintaining user trust and transparency.

Core AI application

Conversational agent uses Bloomreach's proprietary e-commerce-optimized agentic framework to power intelligent shopping conversations. Through Bloomreach’s partnership with Google, the framework runs on the latest Gemini foundation models.

Bloomreach Weblayers embed the Conversational agent experience on a website and display relevant personalized content at the right time and location for each shopper.

Query processing

Conversational agent uses a powerful agentic architecture to handle user queries. It connects with Bloomreach's extensive suite of tools for tasks ranging from product catalog searches (via either Search or Conversational agent Search) to personalized responses based on past shopper behavior.

Conversational agent has an optional capability of using LLMs to augment product feeds by improving descriptions and generating new product attributes.

Secure data handling

Data architecture

Conversational agent ingests, processes, and secures three types of data:

• Product data: Conversational agent ingests your product data either via Conversational agent Search or Search. This allows Conversational agent to truly understand your product feed and find matching products not just through exact matches, but by grasping their deep semantic meaning.

• Customer data: Conversational agent uses Bloomreach's data engine to store and access information about shoppers. For existing Marketing customers, all historic data stored in the data engine is available for use within Conversational agent.

• General knowledge data: Conversational agent stores general knowledge data, which includes textual data. For instance, PDFs, raw text, or URLs.

Check this article on data captured and stored by Conversational agent.

Security and privacy measures

To ensure privacy and data security, Conversational agent:

• Runs a private instance for all calls to Gemini LLM models.
• Never shares customer data with third parties to train the underlying models.

Robust AI risk management system

Guardrails module

Conversational agent mitigates risks via runtime guardrails that limit any potentially malicious requests:

• Run input guardrails to block malicious user behaviour. Conversational agent maintains the context and responds to malicious user requests with "Sorry, I can't help you with that".
• Run response guardrails to ensure Conversational agent’s responses are safe, block potentially unsafe responses, and match the guidelines defined within your tone of voice.

Safety testing

Bloomreach performs regular and rigorous testing to ensure safe Conversational agent responses. The testing ensures that Conversational agent follows proper business processes, protects confidential business information, prevents unauthorized access, and keeps conversations appropriate and relevant.

Conversational agent's guardrails, along with its agentic framework, prevent it from generating "unsafe" responses. This includes:

• Answering questions irrelevant to the business context (for example, responding to "Who is the US president?").
• Discussing topics prohibited by the Bloomreach customer (for example, politics, war).
• Using terminology prohibited by the Bloomreach customer (for example, the term "hand-made").

Handling of AI bias

Conversational agent, by design, minimizes AI bias by referencing product data and limiting the use of raw LLM knowledge.

Product feed reference

Conversational agent uses your product feed to answer product-related queries. Unless the product feed is outdated or incorrect, Conversational agent consistently presents accurate and relevant information sourced directly from the product feed, reducing risks related to biased responses.

Selective use of LLM knowledge

Conversational agent leverages your product data as the primary source to answer product-specific questions. General LLM knowledge serves as supplementary reasoning logic only when relevant. This reduces misinformation risks.

For example, if a shopper asks, “Will this mattress fit my bed with dimensions 76x80 inches?”, Conversational agent verifies using the product feed and comparative reasoning to check whether the product matches the given dimensions.

How Conversational agent limits use of raw LLM knowledge

Conversational agent uses the input and output guardrails to filter relevant queries and check responses based on the guidelines you set.

Conversational agent’s agentic framework uses the LLM as an intelligent orchestrator that selects appropriate tools and actions. The LLM's role is to:

• Identify which Bloomreach tools are relevant for a query.
• Determine the sequence of actions, such as performing a search or checking an FAQ article using RAG (Retrieval-Augmented Generation).
• Present results in a conversational format.

This architecture ensures that the LLM coordinates tools rather than relying on raw knowledge. However, as with all LLM-powered tools, errors can occur.

External sources of bias

Conversational agent may inherit bias from external sources it can't control:

• Third-party AI models, including LLM and embeddings from Google.
• Merchant-provided data, including product catalogs and Search merchandising rules.

The platform itself doesn't introduce additional bias through its agentic orchestration logic.

AI use transparency

The Conversational agent chat widget displays a customizable AI disclaimer at the bottom to ensure transparency regarding AI use. The default disclaimer is “Responses use AI and may contain errors.”

Disclaimer customization

The Bloomreach team can edit the default AI disclaimer to match your legal constraints. Contact your Bloomreach representative to modify the disclaimer.

Disclaimer truncation

To ensure the disclaimer appears on a single line, longer disclaimers are truncated using a fixed character cutoff, rather than a dynamic cutoff based on the screen size. Any text exceeding the cutoff is hidden behind a tooltip.

Cutoff lengths

• Mobile: 50 characters
• Desktop: 65 characters

AI compliance at Bloomreach

To learn more about the compliant development and deployment of AI at Bloomreach, contact:

Data protection officer: [email protected]

At Bloomreach, we're deeply committed to building AI the right way. Visit this article on Bloomreach’s AI development principles.