Test Search
The test search panel lets you run semantic search queries against a table directly — without going through the AI conversation. Use it to verify that rows are retrieved correctly and that your search settings produce relevant results before publishing the agent.
Navigate to Actions → Plugins → Dynamic Table → Tables, open the table you want to test, and go to the Test Search tab.
Running a Search
Enter a natural language query in the Query field and click Search.
Example queries:
laptop under 3kg with USB-Csenior engineer based in Kuala Lumpurhalal certified products from local supplier
The plugin converts the query into a vector embedding and runs a similarity search against all rows in the table. Results are returned based on the search settings configured for this table.
Overriding the Similarity Threshold
You can enter a custom threshold in the Threshold field to override the table's default similarity threshold for this test run only. This does not change the saved configuration.
Use this to quickly experiment with different threshold values before committing to a change in Search Settings.
Reading the Results
Each result shows the matched row's column values. If Return Similarity Score is enabled in search settings, each result also includes a score between 0 and 1.
| Score range | Interpretation |
|---|---|
0.85 – 1.0 | Very strong match — the row is semantically very close to the query. |
0.70 – 0.84 | Good match — expected results at the default threshold. |
0.50 – 0.69 | Moderate match — may be retrieved if tolerance increment is enabled. |
< 0.50 | Weak match — unlikely to be useful without lowering the threshold significantly. |
Interpreting No Results
If the search returns no rows:
- Check that the table is active. Inactive tables are excluded from search.
- Check that rows exist. An empty table returns nothing.
- Check that at least one column is marked as Searchable. If no columns are searchable, all row embeddings are empty.
- Lower the threshold. Enter a lower override threshold in the test panel and run again.
- Check if embeddings are up to date. If searchable columns were added or changed after rows were imported, regenerate embeddings.
Improving Search Quality
Problem: Results are returned but not the right ones
The similarity threshold may be too low, allowing weakly related rows in.
- Raise the threshold to
0.8or higher and test again. - Check the similarity score of the wrong results — if they are below
0.7, raise the base threshold.
Problem: The right row is not returned
The row's embedding may not represent the query well enough.
- Check that the relevant columns are marked as Searchable in the column settings.
- Increase the Search Weight of high-signal columns.
- Add more descriptive values to the row's searchable columns.
- If the column configuration changed recently, regenerate embeddings.
Problem: Results vary significantly by minor query wording
- Enable the LLM Reranker to post-process results using an AI model.
- Add better column descriptions to improve embedding quality.
What the AI Sees
The test search panel shows raw rows exactly as the plugin returns them to the AI. The AI then formats and presents these rows to the user based on the function tool settings (response prompt, rules, and description).
If the raw results look correct but the AI's reply to users seems off, adjust the Response Prompt or Rules in Function Tool Settings.
Related Pages
- Search Settings — tune threshold, tolerance increment, and reranker.
- Create a Table — adjust searchable columns and search weights.
- Function Tool Settings — control how the AI presents search results.
- Dynamic Table Overview