Generative Model

Sends interaction history, contextual variables, and priors to an external large language model (LLM) via a compatible API (for example Groq or OpenAI). The runtime parses JSON scores returned by the model and uses them like other dynamic scores.

Algorithm

Config value: "approach": "behaviorAlgos", "sub_approach": "generative"

1. Build a prompt from the current entity context, historical interactions, and configured priors.
2. Call the external LLM with sampling parameters (for example temperature).
3. Parse structured JSON with per-offer or per-arm scores from the response.
4. Feed parsed scores into the same ranking path as other behavioral algorithms.

Stochasticity and ranking behavior depend on the model, prompt, and temperature.

Parameters

• temperature (inside prompt_parameters or equivalent): Controls randomness of the LLM; default 1.0. Lower values tend to be more deterministic; higher values increase variety.
• Processing Window / Historical Count: Bound how much history is included in prompts when configured on the deployment.

Cold Start

Recommendations are always returned. The real-time training path always produces a scored options array:

• No history: Every offer in the options store receives a uniform random score. The prompt sent to the LLM contains empty or minimal interaction context.
• With history: The LLM receives interaction data and contextual variables, and returns per-offer JSON scores. The quality of scores depends on the model, prompt, and richness of the logs.
• Offers not scored by the LLM receive a random fallback score and are still included in the result.

The scored options are then sorted by arm_reward and handed to the configured dynamic post-score class, which controls the final offer selection and response formatting.

When To Use

• Complex offer selection where natural-language reasoning or unstructured context helps
• Rich unstructured context (notes, policies, descriptions) that is expensive to hand-engineer into features
• Experimental or hybrid setups combining LLM judgment with the ecosystem.Ai pipeline

When NOT To Use

• When you need deterministic, auditable scores on every request
• Latency-sensitive production paths (each score may require a network round trip)
• Cost-sensitive environments where per-request LLM calls are prohibitive

Example

from prediction.apis import deployment_management as dm
from prediction.apis import online_learning_management as ol
from prediction import jwt_access
 
auth = jwt_access.Authenticate("http://localhost:3001/api", ecosystem_username, ecosystem_password)
 
deployment_id = "demo-generative-model"
 
online_learning_uuid = ol.create_online_learning(
        auth,
        algorithm="ecosystem_rewards",
        name=deployment_id,
        description="Generative Model (LLM) configuration",
        feature_store_collection="set_up_features",
        feature_store_database="my_mongo_database",
        options_store_database="my_mongo_database",
        options_store_collection="demo-deployment_options",
        randomisation_processing_count=5000,
        randomisation_processing_window=604800000,
        contextual_variables_offer_key="offer",
        create_options_index=True,
        create_covering_index=True
)
 
online_learning = dm.define_deployment_multi_armed_bandit(epsilon=0, dynamic_interaction_uuid=online_learning_uuid)
 
parameter_access = dm.define_deployment_parameter_access(
    auth,
    lookup_key="customer_id",
    lookup_type="string",
    database="my_mongo_database",
    table_collection="customer_feature_store",
    datasource="mongodb"
)
 
deployment_step = dm.create_deployment(
    auth,
    project_id="demo-project",
    deployment_id=deployment_id,
    description="Generative Model demo deployment",
    version="001",
    plugin_post_score_class="PlatformDynamicEngagement.java",
    plugin_pre_score_class="PreScoreDynamic.java",
    scoring_engine_path_dev="http://localhost:8091",
    mongo_connect=f"mongodb://{mongo_user}:{mongo_password}@localhost:54445/?authSource=admin",
    parameter_access=parameter_access,
    multi_armed_bandit=online_learning
)