| name | optimize-cortex-search-service |
| description | "Use for optimizing existing Cortex Search services. |
Optimize Cortex Search Service Skill
When to Use
When a user wants to optimize an existing Cortex Search service.
Prerequisites
- Fully qualified Cortex Search service name (i.e. DATABASE.SCHEMA.SEARCH_SERVICE)
- Snowflake access configured
Step 1: Get Queries to Run on Search Service
Ask the user for a list of queries to evaluate the Cortex Search service on and then format it into a JSON file of the form:
{
"queries": [
{"query": "search query text"},
...
]
}
If they do not have such a list, generate a list of 30 queries via generate_synthetic_queries.py (which will have the proper formatting) via a command such as:
python generate_synthetic_queries.py \
--service <DATABASE.SCHEMA.SERVICE_NAME> \
--connection <CONNECTION> \
--num-queries 30 \
--output <QUERY_OUTPUT_FILE>
Step 2: Find Optimal Search Weights
Use optimize_search_weights.py to generate the optimal weights for the Cortex Search service via a command such as:
python optimize_search_weights.py \
--service <DATABASE.SCHEMA.SERVICE_NAME> \
--queries <QUERY_FILE> \
--score-unjudged \
--cache-file <CACHE_FILE> \
--n-trials 20 \
--connection <CONNECTION>
The <QUERY_FILE> should be the file generated from the previous step and the <CACHE_FILE> is used to speed up computation of this script within a single run of it and across runs.
Step 3: Recommend Changes to User
Give the user the recommended weights computed in the previous step and suggest they use scoring profiles to optimize their Cortex Search service, but stress that this currently cannot be done for Cortex Search services called by a Cortex Agent.
