How did you implement the scoring mechanism that uses multiple weighted signals such as source type, document version, and user access scope to improve relevance? Did you use any specific frameworks or libraries, and how did you ensure the weights were appropriately calibrated?

I built the scoring logic on top of Elasticsearch’s function_score API, which allows combining multiple factors into a single relevance score. Each signal—source type, document version, and user access scope—was mapped to a separate function. For example, source type was encoded as a categorical field and scored using a field_value_factor; document version was normalized to a numeric score using a script that penalizes older versions; user access scope was represented as a boolean flag that boosts documents the user is allowed to see. The overall score was a weighted sum of these functions, with the weights stored in a JSON configuration that could be updated without re‑indexing. To calibrate the weights, I used a small training set of labeled queries and applied a grid search over weight combinations, evaluating each with precision@k and mean reciprocal rank. I also incorporated a Bayesian optimization loop to fine‑tune the weights in production, feeding back click‑through data. For the implementation, I leveraged the Elastic LTR (Learning to Rank) plugin to train a linear model that learns the optimal weights from historical search logs, ensuring that the system adapts to changing user behavior over time.