Guidelines for slot filling in catalog request
Summarize
Summary of Guidelines for slot filling in catalog request
Now Assist in ServiceNow uses a large language model (LLM) to perform slot filling by extracting variable values from a requester’s conversational input and pre-filling catalog item questions. This reduces the manual effort required by requesters to complete catalog requests. Slot filling leverages the LLM to identify relevant questions and extract matching answers, enhancing the efficiency of catalog requests within the conversational interface.
Show less
Because LLMs operate probabilistically, slot fill accuracy can vary and cannot be guaranteed, but the objective is to maximize pre-fill accuracy as much as possible.
Catalog item question configuration
Effective slot filling depends heavily on how catalog item questions are configured. To improve accuracy:
- Use concise, clear, and unambiguous question labels in the Question field to help the LLM correctly identify what information to extract.
- Avoid complex, multi-part labels; prefer single focused questions (e.g., “What issue are you experiencing?” instead of multiple combined queries).
- Differentiate question names clearly in the Name field to prevent inconsistent filling, especially when similar questions exist (e.g., distinguish “Requested For” from “On behalf of”).
- Adopt a naming convention where the object leads and the attribute follows (e.g., daterequired rather than requireddate) to help the LLM match inputs accurately.
- Replace system-generated or technical question names with human-readable, descriptive names that reflect the actual item (e.g., use Laptopcharger instead of Hardware001P04power).
- Check and maintain the
com.glide.cs.genai.discovery.limits.skill.slotsproperty at its default value to ensure the maximum number of questions are sent for slot filling. Reducing this limit will restrict pre-filling capacity.
Model selection
The quality of slot filling is closely tied to the capability of the LLM model configured in the instance’s Conversational Catalog skill:
- Older or smaller LLM models may struggle with ambiguous inputs, multi-variable extractions, and resolving references such as duplicate user names.
- More recent and capable models improve the extraction of complex expressions like dates (“next Tuesday,” “in two weeks”) and can handle multiple variables from a single input more effectively.
- To verify or change the model, navigate to Now Assist Admin > Skills and review the model assigned to the Conversational Catalog skill.
- For guidance on available models suited to your entitlement, contact your ServiceNow account team.
Now Assist in Conversational Catalog Request uses a large language model (LLM) to extract variable values from a requester's input and pre-fill catalog item questions, a capability known as slot filling.
When a requester describes their need in the conversational interface, Now Assist passes that input to an LLM, which identifies the relevant catalog item questions and extracts matching values from the requester's message. Those values are then pre-filled into the catalog item request form, reducing the number of fields the requester must complete manually.
Because LLMs are probabilistic by design, slot fill accuracy can vary across sessions even for identical inputs. Slot fill accuracy cannot be guaranteed. The goal is to maximize pre-fill accuracy.
Catalog item question configuration
The LLM uses question labels and descriptions to determine what information to extract from the requester's input. Ambiguous or poorly written question labels are a leading cause of missed or incorrect slot fill. The following considerations apply when configuring catalog item questions for use with slot filling:
- Keep the Question field label concise and unambiguous. Avoid long or multi-part labels. For example, use a single focused label such as "What issue are you experiencing?" rather than combining multiple questions into one label.
- Avoid overlapping question names in the Name field. When two questions in the same catalog item have similar labels, for example "Requested For" and "On behalf of", the LLM may fill them inconsistently. Differentiate question names clearly to reduce ambiguity.
- In the Name field definition, lead with the object and follow with the attribute. For example, use
date_requiredrather thanrequired_date. This structure enables the LLM to identify the correct question. - Replace system-generated prefixes and internal question names with human-readable names. For example, replace
Hardware_001_P04_powerwithLaptop_charger. - Use names that reflect what the item actually is, so the LLM can match requester input to the correct question reliably.
- Check the value of the com.glide.cs.genai.discovery.limits.skill.slots property, which controls how many questions are sent for slot fill. If the value is set to a value less than the default, questions beyond that limit aren't pre-filled. Restore the value to the default unless there is a specific reason to reduce it.
Model selection
Slot fill quality is directly influenced by the capability of the LLM configured for the instance. Older or smaller models have reduced ability to interpret ambiguous input, resolve references such as matching a name to a user record when duplicates exist, and pre-fill multi-variable catalog items.
If slot fill failures persist on catalog items with clear, well-configured question labels, check which model is configured for the instance's Conversational Catalog skill. A more capable or recent model typically improves the following:
- Extraction of date and relative time expressions, for example "next Tuesday" or "in two weeks".
- Multi-variable extraction from a single requester utterance.
To review the model configuration, navigate to and check the model assigned to the Conversational Catalog skill. Contact your account team for guidance on available model options for your entitlement.