DeepSearch + GenSearch Engine Settings

This unique model combination uses the strenths of our DeepSearch Model alongside our GenSearch Model to provide a fast response for short-form queries and an accurate response for longer-form queries.  The interplay between the two is controlled by a trigger that can be customized by you.

GenSearch Trigger

When composing an Engine, you want to think with the practical use that the Engine will be accommodating. You can switch the engine from DeepSearch to GenSearch by using a word count or a character count. If the Trigger is set to "Word Count" and the Threshold is set to "3", if the user types 2 words, the request will be sent to DeepSearch, and anything after 3 words will be sent to GenSearch.

Example: the setup for a Search Bar could be to use the stepped approach of DeepSearch + GenSearch, with a word count of “2” as a trigger.

The output would then be:

• Any query of one word would be sent to DeepSearch

• Any query of two words or more would be sent to GenSearch

It is easy (and important) to test all variations within the Sandbox to find the best configuration for your use case. Remember that the number of Engines you can compose is unlimited, allowing you to dial in Engines that work best for any number of use cases.

Query Settings

Query Settings are options that can be enabled to augment the query before it hits the models that you have selected.

Spell Correction

The XSearch spelling correction plugin compares the query to a comprehensive language library (that is custom-trained on your individual product dataset) to fix spelling mistakes before running the search.

When to use Spell Correction

Both DeepSearch and GenSearch have built-in features to accommodate for spelling issues and typos automatically, and the Spell Correction plugin should only be used as needed. DeepSearch has fuzzy searching (which accommodates for deviations in copy), and GenSearch has built-in LLM understanding of the query regardless of spelling mistakes.

Negative Tone

If you find that users are using negatives in their searches (“Dress that is not white”), you’ll find that traditional search methodology doesn’t understand how to process this, and will ignore the negative term (“not”). Turning this feature on will use AI to adjust the query automatically and help the Engine respect the negative and provide more relevant results in these use cases.

Explicit Search

If your use case includes the possibility of searching by SKU or exact product name, Enhanced Explicit Search will do a literal lookup of the query against SKU and Product Name before running the search through DeepSearch or GenSearch, ensuring that these specific use cases are properly handled 100% of the time. Note: This will only work for exact matches, this feature will not respect deviation.

Query Guard

Due to the open and interpretative nature of GenSearch, it is technically possible to return search results for inappropriate queries (political topics, illegal topics, and other potentially controversial subjects). To solve for this, Query Guard can be enabled to restrict or adjust any such queries.

Query Guard operates at either the subject or keyword level, and you have the ability to:

• Add specific keywords to redact

• Select from a list of pre-defined subjects to avoid

• Create your own subject to avoid

• Add keywords to never-alter (if there is a word that is in a product title or IS relevant to your product line, this allows you to bypass Query Guard for any related terms)

Once you have configured what to avoid, you can then determine what is done with queries that are identified:

• Return zero search results

• Zero search results PLUS a text response

• Edit the query to remove the flagged topic

DeepSearch Settings

Customize the result-settings by limiting the number of products returned in any query, or adjusting the relevancy of a result to a query’s intention.

DeepSearch Product Return Limit

While the engine will explore the entire catalog for your search, the Product Return Limit controls how many products are displayed to customers. Reference application and website traffic analytics to determine where your customers drop-off in their discovery journey to inform this figure.

DeepSearch Relevancy Threshold

At its most fundamental level, DeepSearch ranks every product in the catalog with a relevancy score against the query, with “1” being a perfect match, and “0” being totally irrelevant. The Relevancy Threshold allows you configure how “open-ended” you want the search results to be, with a high value being restrictive and a low-value allowing it to return products that might be less relevant.

DeepSearch Weighting

Blending two different types of search methods, the weighting section helps you find the right combination for your catalog and customers. The final relevancy output will be a combination of both Neural and Keyword searches, so the higher the value you put on each, the stronger impact they will have respectively.

Best Practice

The best practice is to keep Keyword Search at .05 or .1, to allow the Neural Search to provide the majority of the results.

Neural Search Weight

Neural search is a vectorized, concept-based search methodology that has a strength range from 0-1.

Keyword Search Weight

This weighting provides control over a more conventional search tool that uses literal, keyword search methodology. This can be informed by your application or web site analytics, and depends on how your customers search.

Keyword Search Fuzziness

Fuzziness allows your search engine to look for other similar terms or words you are asking for instead of rigid keyword to keyword exact match.

GenSearch Settings

If your semantic behavior configuration has GenSearch in the pipeline, all of the GenSearch specific configuration will be available.

Generative LLM Settings

There are two types of LLMs that can be leveraged within GenSearch:

• Accurized (standard, highly accurate): This is the standard LLM that should be used, so long as you plan to leverage the caching system.

• Accelerated (faster, lower accuracy): If you plan to avoid caching entirely, this model will allow you to maintain a faster response time, though the accuracy will be slightly lower.

GenSearch Product Price Enhancement

If you want to enable queries that are directly correlated to the price data point in your product catalog, such as “Jeans under $100” (price), then you can activate the result enhancement switch here.

GenSearch Thresholds

At its most fundamental level, GenSearch ranks every product in the catalog with a relevancy score against the query, with “1” being a perfect match, and “0” being totally irrelevant. The Relevancy Threshold allows you configure how “open-ended” you want the search results to be, with a high value being restrictive and a low-value allowing it to return products that might be less relevant.

GenSearch Translation

GenSearch has a built-in LLM-powered translation service that enables all queries to work in all languages, with an auto-detect to determine the language of the query itself. If you need to override this and manually set the input language for improve accuracy, you can do that here.

Response Settings

Response Settings allow you to configure what happens after a query has been responded to.

Disable Cache

Enable Cache

Caching is an important part of maintaining speed and performance, and you have the ability to fully control how caching is set up at the Engine level.

It is important to understand that the cache exists at the query level, meaning that each individual query is itself cached until either the TTL is hit (from 6 hours up to 30 days, based on configuration) or there is a change in the product catalog from an update.

Sorting

Sorting allows you to determine how the results will be sorted, after the Engine has isolated the relevant products using all previous configuration. The default is by relevancy to the query itself, however this can be overridden with the following algorithms:

Sorting Options