The Match Maven module enables data teams to build and evaluate match-merge models using large language models (LLMs) and embedding-based similarity techniques. It is designed for experimentation, iteration, and optimization of custom entity resolution models that support advanced MDM use cases on Databricks.

Step 1: Model Creation

• Access Match Maven via the dedicated navigation card on the LakeFusion interface.

• Select the Target Entity you want to experiment on (e.g., Legal Entity, Customer, Product).

• Configure the Required Parameters:

• Model Name: Enter a unique identifier following naming conventions.

• Description: Provide a detailed explanation of the model’s purpose and scope.

• LLM Model Selection: Choose from supported large language models.

• Embedding Model Specification: Define the embedding model that will generate semantic representations.

• Attribute Selection: Select key entity attributes relevant to the match logic.

• Select model source from available options:

• Databricks foundation models

• Databricks custom implementations

• External Model API integrations

• Execute Model Creation by clicking the Add button.

New Feature - Info Button

An Info Button has been added to each experiment entry. Use this to view detailed metadata and insights on the model’s structure, run statistics, and version information.

Step 2: Experiment Execution

• Navigate through the four sequential stages:

Once a model is defined, run a guided experiment through four sequential stages:

• Stage 1 - Prepare

• Launch the experiment using the Run Experiment option.

• Match Maven will preprocess and organize the input dataset.

• Stage 2 - Experiment

• The experiment typically runs for approximately 40 minutes, depending on model complexity and data volume.

• Real-time progress can be monitored through the dashboard.

• During this stage, you can apply custom prompting techniques to influence how the LLM interprets specific data contexts. This helps align the model’s semantic understanding with the business logic behind your matching use case.

• Stage 3 - Review

• Upon completion, review generated match pairs and evaluate the model’s effectiveness using embedded accuracy and confidence metrics.

• You can inspect match results at the record level, compare scores, and analyze the logic driving match decisions.

• Stage 4 - Approved

•    If the experiment yields satisfactory results, approve the model to move forward to threshold configuration.

Step 3: Threshold Configuration

• Once a model is approved:

1. Define Merge Thresholds to automatically combine high-confidence matches.

2. Set Match Thresholds for records requiring manual review.

3. Establish Non-Match Thresholds to exclude low-confidence pairings from further consideration.

When thresholds are configured and the model is published, notifications are sent to the designated business owner with a prompt to review flagged critical records.

Step 4: Review Record

After publishing, flagged records are accessible via the Entity Search or Notifications Dashboard.

• Click on each flagged match to view comprehensive attribute comparisons, source system information, and scoring details.

• Evaluate each record to determine whether to merge, exclude, or escalate for further stewardship.