Search Settings

The Dynamic Table plugin uses vector-based semantic search to retrieve rows. Search settings control how similarity is measured, how far the system will lower the bar to find results, and whether an LLM reranker should refine them afterward.

Each table (collection) has its own search settings. Navigate to Actions → Plugins → Dynamic Table → Tables, open the table, and go to the Search Settings tab.

Similarity Threshold

The similarity threshold determines the minimum score a row must reach to be included in search results. Scores range from 0 to 1, where 1 is an exact semantic match.

Default: 0.7

Start with the default and adjust based on test search results. If the agent returns too many irrelevant rows, raise the threshold. If it misses rows that should match, lower it.

Tolerance Increment

Tolerance increment automatically lowers the similarity threshold in steps when the search returns fewer results than the configured minimum. This prevents the agent from returning an empty response when a strict threshold yields nothing.

Tolerance Increment Settings

How It Works

1. Search runs at the starting threshold.
2. If the number of results is below the minimum results setting, the threshold is reduced by the threshold step.
3. The search runs again at the lower threshold.
4. This repeats until either the minimum results are reached, the maximum results cap is hit, or the threshold reaches the minimum threshold.

Example with defaults:

Set Minimum Results to 0 to disable tolerance increment and always return results at the base threshold only.

Similarity Score in Results

Enable Return Similarity Score to include the numeric similarity score alongside each result row. This is useful during testing to understand how well rows are matching, but can be omitted in production if you do not want scores visible to the AI.

Default: disabled

LLM Reranker

The reranker is an optional second stage that uses a language model to re-score and reorder results after the vector search. It reads the original user query and the retrieved rows and returns them sorted by relevance.

Use the reranker when:

• The table contains rows with similar wording but different meaning.
• Precision matters more than latency.
• Vector similarity alone is not producing accurate enough results.

Reranker Settings

Enabling the reranker adds latency because it requires an additional LLM call after the vector search. Enable it only for tables where search precision is critical.

Regenerating Embeddings

Each row's vector embedding is generated when the row is created or updated. If you change which columns are marked as Searchable, or adjust Search Weight values, existing rows retain their old embeddings and will not reflect the updated configuration.

To rebuild all embeddings for a table:

1. Open the table in the Tables tab.
2. Click Regenerate Embeddings.
3. Confirm the operation.

The system processes rows in batches and shows a progress indicator. The table remains searchable during regeneration, but results may be mixed until the process completes.

Regeneration consumes OpenAI API credits proportional to the number of rows and the size of each row's text content.

When to Regenerate

Related Pages

• Create a Table — configure column search weights and searchable flags.
• Test Search — validate search quality after changing these settings.
• Dynamic Table Overview