Conversational Intent-Based Recommenders

Why Conversations Need a Different Signal

Web personalization relies on implicit signals, such as clicks, views, and time on page. Conversations work differently.

• Users state what they want in plain language
• Intent can span multiple turns and use informal or ambiguous language
• The user's current ask outweighs historical behavior when the two conflict

For example, a shopper buying a gift doesn't want recommendations based on personal preferences.

To handle this difference, DLPR treats conversational intent as a primary engagement signal and consumes it in real time.

How a Conversational Recommender Works

A conversational recommender has two distinct flows - a training flow and an inference flow. Training runs offline on a regular cadence and determines whether the recommender serves DLPR-based or semantic-search-based recommendations. Inference runs in real time during each Agentforce conversation.

Training Flow

Training prepares the recommender and determines whether it operates in cold start. The recommender evaluates whether the engagement data is robust enough to power a DLPR model. When the data meets this standard, the recommender activates DLPR. Otherwise, the recommender uses semantic search to deliver recommendations.

Inference Flow

Inference runs every time a user interacts with the agent.

1. Initiate the agent chat: The Agentforce agent fetches user details using hidden prechat variables.
2. Classify the utterance: The agent classifies the user's utterance into the Personalized Recommendations subagent and triggers the subagent's actions in sequence.
3. Extract user intent: The Agentforce action Understand User Intent uses an LLM to extract a concise, structured intent from the user's utterances. This action bridges vocabulary gaps between user utterance and your catalog and supports multiple language translations.
4. Get recommendations: The agent invokes Get Recommendations, which calls the personalization decisioning pipeline. The recommendation engine routes the request based on the recommender's training outcome.
5. Return recommendations: If the recommender is in cold start, the agent receives semantic-model-based recommendations. If not, the agent receives DLPR plus text-embedder-based recommendations. The agent presents the recommendations inside the conversation.

Cold Start in Conversational Recommenders

A conversational recommender determines cold start at training time, not at request time. If DLPR's evaluation metric on the test split doesn't beat the semantic search baseline, the recommender stays in cold start and serves semantic-search-based recommendations. As engagement data accumulates and the next training run produces a stronger DLPR metric, the recommender exits cold start and switches to DLPR plus text-embedder-based recommendations.

Model Inputs

• Live conversation: The user's utterances across the current session.
• Catalog metadata: Item titles, descriptions, categories, and attributes, embedded for semantic matching.
• Engagement history: Past interactions when available.

Guidance for Your Catalog

Because intent is matched semantically to catalog items, the richness of your catalog descriptions directly determines recommendation quality.

• Write complete titles and descriptions that reflect the terms users naturally ask for.
• Include category and attribute metadata for every item.
• Keep metadata current as your catalog evolves.