VWAM Help

Appearance

AI providers & how prompts are handled ​

Two external AI providers supply the large language models behind the VWAM assistant. VWAM uses them as sub-processors under contract. This page explains which AI providers VWAM uses, what VWAM sends to them, and how their data retention works.

Which AI providers ​

VWAM uses OpenAI and Anthropic through their standard APIs. VWAM splits the work by task. In the standard configuration, OpenAI models generate the main conversational responses and tool-driven reasoning, and handle routing, classification, summarization, reranking, brand rendering, and vision extraction. Anthropic models handle general attribute and property extraction.

Each task also has a fallback on the other AI provider. If the primary AI provider is unavailable, the other one serves the request. Both AI providers therefore process request content in normal operation and in failover. VWAM can adjust the task-to-provider assignment per contract.

Product search and knowledge-base search use OpenAI text embeddings. VWAM sends the searchable content to OpenAI at import time, which covers product catalog records and knowledge-base documents. VWAM also sends the search queries that it derives from the conversation. OpenAI returns the vectors that back the search.

What VWAM sends to an AI provider ​

Each request to an AI provider includes:

  • The system prompt — the assistant's instructions. When a shopper is on a product page, the prompt also carries the details of that product.
  • The recent conversation history — the shopper's messages and the assistant's replies. The platform also injects structured product context into this history, such as the titles, prices, and identifiers of the products that it shows or references.
  • A running summary of older turns, once a conversation grows long.

A merchant can connect a customer data platform (CDP), for example BlueConic or Klaviyo. Selected visitor-profile attributes can then reach the system prompt for personalization. The merchant controls which attributes VWAM sends. An attribute reaches the model only if the merchant maps it into a per-property allowlist, or enables identity personalization. Identity personalization surfaces the standard identity fields. VWAM sends nothing from a connected CDP by default.

Retention and model training ​

The AI providers never use shopper conversations to train their models.

Each AI provider has its own retention model.

  • OpenAI — every chat-model request that VWAM sends opts out of OpenAI's server-side request retention, on all reasoning and tool-calling paths. Our OpenAI account and our data-processing terms govern actual retention, not the per-request flag alone.
  • Anthropic — the Anthropic API exposes no per-request retention flag. Account-level settings, organization-level settings, and our agreement with Anthropic govern retention there.

Contact us for our current sub-processor list and for data processing agreement (DPA) detail.

Storage of responses ​

VWAM stores the assistant's replies and the conversation history in its own database, to keep conversation state across turns. This storage is platform-side, and it is independent of the AI providers. The same encryption, isolation, and data-subject-request handling apply as for the rest of your conversation data.