Configure AI Prompt for Copilot for Agents KB Query Extraction
You can use NiCE CXone GenAI Prompts to control the prompt that Copilot for Agents uses to perform KB query extraction. This lets you guide how Copilot for Agents interprets the conversation and formulates knowledge base search queries, including the level of detail, language, and any special instructions you want applied.
There are two KB Query Extraction categories:
-
KB Query Extraction (Voice): For voice interactions. This category was called KB Query Extraction before. Your existing prompt profiles are in it and work the same way.
-
KB Query Extraction (Digital): For digital interactions. The prompt reads each message the contact sends, decides whether the knowledge base can answer it, and rewrites it as a clear search question. Copilot for Agents then sends that question to the knowledge base instead of the contact's whole message.
Prompt profiles in one category do not appear in the other.
Once you have completed your prompt configuration, you can import or enter up to five transcripts to test the settings against. Copilot for Agents uses your custom prompt to extract KB queries from each transcript so you can preview how well the configuration surfaces the right intents and topics.
If you are not satisfied with the queries generated in the test, you can continue to refine your settings and prompt instructions until you are satisfied. After that, you can publish the prompt to make it available for selection within the Copilot for Agents profile on the Copilot for Agents Config page.
-
All users can now open the LLM list. Which models you can select depends on your organization's licensing. The list itself is grouped by standard and premium models, with models verified by NiCE CXone shown first and the rest listed alphabetically.
-
If you do not have the Prompt Editor Advanced license, premium models appear in the list but you cannot select them.
-
To see which models are premium, open the LLM Manager.
-
Premium models are billed based on usage. Your organization is charged for the number of tokens the model reads and writes when a prompt runs. To use premium models, your organization needs the LLM consumption item on its contract. Contact your NiCE CXone Account Representative for pricing.
-
NiCE Default is available as an LLM option. It is the model that NiCE CXone sets for that specific prompt type. If a prompt template only supports one model, NiCE Default is the only option shown and the LLM list is disabled. If a template supports multiple models, NiCE Default appears alongside them as a selectable choice.
-
If the LLM currently selected for a profile is scheduled for retirement, NiCE CXone displays a warning next to the LLM list stating that the model will retire in a set number of days and prompting you to update to a supported model.
-
When a model you use is closer to its retirement date, NiCE CXone also displays an LLM Update dialog when you open the affected profile. The dialog states the date the model switches to retention mode, after which the model is no longer accessible, and reminds you to update the LLM and test and publish all affected profiles before that date. Select Don't show this message again to suppress the dialog for that profile, or click OK to close it.
-
If a model reaches its retirement date and you have not selected a different model, NiCE CXone automatically switches the profile to NiCE Default. Test and publish the profile after this switch to confirm the output still meets your requirements.
Each prompt profile in GenAI Prompts uses a large language model (LLM) to generate its output. You select the LLM in the LLM list at the top of the profile.
The list groups models into Standard and Premium. NiCE-verified models appear first in each group, followed by the remaining models in alphabetical order. All users can view the list. Which models you can select depends on your organization's licensing.
-
These tables list only the LLMs that NiCE CXone has verified. Verified models differ by prompt type, so check the Verified LLMs by Prompt Type table before you change the model in a profile. The LLM list in the application may show additional models that have not yet been verified by NiCE CXone. In the list, the
icon next to a model indicates it has been verified by NiCE CXone for that prompt type. You can select any listed model, but if you choose one that is not listed for your prompt type, testing that model for your specific use case is your responsibility. -
NiCE Default is the model that NiCE CXone sets for each prompt type. NiCE CXone chooses it by testing, and the model can change over time, for example when a prompt is updated or a model is retired. Because the model differs by prompt type and can change, NiCE Default is not listed as a separate row in the LLM Ratings table.
-
Profiles previously defaulting to a specific model remain on the configured model, with no functional change.
This table shows the LLMs you can choose from, along with a rating from 1-10 for the speed, accuracy, complexity handling, and hallucination resistance. 1 is always the lowest and 10 is always the highest score. A high score is desirable for speed, accuracy, and complexity handling, but not for cost and hallucination resistance.
| LLM Model | Speed | Accuracy | Complexity Handling | Hallucination resistance | notes |
|---|---|---|---|---|---|
| Nova Micro | 10 | 6 | 5 | 6 | Lightweight, low latency; limited reasoning. |
|
Nova Lite |
9 |
7 | 6 | 7 | Higher intelligence than Micro; still lightweight and fast. |
| Nova Pro | 8 | 8 | 8 | 8 | Good balance of intelligence, reasoning, and speed. |
| Claude 4.5 Haiku | 8 | 8 | 7 | 8 | Strong accuracy and reasoning; suitable for efficiency-focused use cases. |
| Claude 5.5 Haiku | 6 | 8 | 7 | Not yet rated | Summary quality matches Claude 4.5 Haiku. It reasons before answering, so it responds more slowly. |
| Claude 4.5 Sonnet | 7 | 9 | 9 | 9 | Flagship-level intelligence and reasoning. |
| Gemma 3 12B | 9 | 8 | 8 | Not yet rated | Fast and low-cost. Scored close to Claude 4.5 Sonnet on interaction summaries in NiCE CXone testing. Less consistent on self-service and real-time summaries. Not available in all regions. |
| Gemma 3 27B | 8 | 7 | 5 | Not yet rated | Slower and less consistent than Gemma 3 12B, especially on journey summaries. Not more accurate than Gemma 3 12B. |
| GLM 4.7 Flash | 8 | 7 | 6 | Not yet rated |
Usually very fast, but some responses take 3-4 seconds. Mid-range accuracy on interaction summaries. |
| Llama 4 Maverick 17B | 9 | 8 | 6 | Not yet rated | Fast. Less consistent than Llama 4 Scout 17B on long interaction summaries. |
| Llama 4 Scout 17B | 9 | 8 | 8 | Not yet rated | Fast and consistent on interaction summaries, journey summaries, and knowledge answers. |
| Ministral 3 8B | 10 | 7 | 5 | Not yet rated | Very fast. Strong on short extraction tasks. Weaker on journey and real-time summaries. |
| Ministral 3 14B | 9 | 7 | 4 | Not yet rated | Fast and fine for interaction summaries and knowledge answers. Unreliable on journey summaries. Not more accurate than Ministral 3 8B. |
| Mistral Large 3 | 8 | 8 | 6 | Not yet rated | Strong on long interaction summaries. Inconsistent on journey summaries. |
| Nova 2 Lite | 9 | 8 | 7 | Not yet rated | Similar speed to Nova Lite and about as accurate as Nova Pro overall in NiCE CXone testing. Can exceed word limits or answer in the wrong language on some prompts. Not available in all regions. |
| NVIDIA Nemotron 3 Super 120B | 9 | 7 | 7 | Not yet rated | Fast. Limited testing so far. Test with your own inputs before you use it. |
| NVIDIA Nemotron Nano 3 30B | 10 | 6 | 5 | Not yet rated | Fastest model tested, but it failed more summary tests than any other model. Use only for short, simple tasks, and check the output. |
| Qwen3 32B | 9 | 8 | 8 | Not yet rated | Fast and consistent on interaction and journey summaries. Not available in all regions. |
| Qwen3 235B A22B 2507 | 8 | 8 | 8 | Not yet rated | Consistent quality across summary prompts. Slower than Qwen3 32B on long interactions. Not available in all regions. |
| Qwen3 Next 80B A3B | 8 | 8 | 7 | Not yet rated | Good summary quality. Slower than Qwen3 32B. Not available in all regions. |
Ratings are relative and based on NiCE CXone testing of summary and knowledge-answer prompts and measured response times. Availability varies by region, and some models are not available in specific regions. Test with your own inputs before you publish a profile.
These ratings describe each model in general. Check the Verified LLMs by Prompt Type table to see which models are verified for your prompt type.
| Prompt Type | Standard Verified Models | Premium Verified Models |
|---|---|---|
| Automated Summary (AutoSummary) | NiCE Default | None |
|
Copilot for Agents Autosummary |
NiCE Default Gemma 3 12B Gemma 3 27B GLM 4.7 Flash Llama 4 Maverick 17B Llama 4 Scout 17B Ministral 3 14B Mistral Large 3 Nova 2 Lite Nova Lite Nova Pro NVIDIA Nemotron 3 Super 120B NVIDIA Nemotron Nano 3 30B Qwen3 235B A22B 2507 Qwen3 32B Qwen3 Next 80B A3B |
Claude 4.5 Haiku Claude 4.5 Sonnet |
| Journey Summary |
NiCE Default Gemma 3 12B Gemma 3 27B Llama 4 Maverick 17B Nova 2 Lite Nova Lite Nova Micro Nova Pro NVIDIA Nemotron Nano 3 30B Qwen3 235B A22B 2507 Qwen3 32B Qwen3 Next 80B A3B |
Claude 4.5 Haiku |
| KB Answers |
NiCE Default Claude 5.5 Haiku Nova 2 Lite Nova Lite Nova Pro |
Claude 4.5 Haiku |
| KB Query Extraction - Voice |
NiCE Default Ministral 3 8B Nova Lite Nova Micro |
None |
| Real-Time Sentiment |
NiCE Default Gemma 3 12B Nova Lite Nova Micro |
None |
| Real-Time Summary |
NiCE Default Nova Lite Nova Micro |
None |
| Email AutoSummary |
NiCE Default Nova Lite |
Claude 4.5 Sonnet |
| Email Generation |
NiCE Default Nova Micro |
Claude 4.5 Haiku Claude 4.5 Sonnet |
| Email Real-Time Summary |
NiCE Default Nova Micro |
Claude 4.5 Haiku Claude 4.5 Sonnet |
| Email Rephrase |
NiCE Default |
Claude 4.5 Haiku Claude 4.5 Sonnet |
| Email Sentiment |
NiCE Default Nova Micro |
Claude 4.5 Haiku Claude 4.5 Sonnet |
| Email Simplify |
NiCE Default |
Claude 4.5 Haiku Claude 4.5 Sonnet |
| Email Topic Extraction |
NiCE Default Nova Lite |
Claude 4.5 Haiku Claude 4.5 Sonnet |
Consider Another Verified Model:
Consider another verified model when testing shows that NiCE Default does not meet a specific requirement, for example:
-
Languages other than English: Automated Summary (AutoSummary) writes summaries in the language set on the Copilot profile. If summaries come out in the wrong language, first check the profile language and any language instructions in your prompt. Then compare other verified models on your own non-English interactions.
-
Long or complex interactions: Choose a model with a high Complexity Handling score, and compare the output on your longest interactions.
-
Real-time prompt types: Real-Time Summary and Real-Time Sentiment run while the interaction is in progress, so response time matters. Choose a model with a high Speed score, and also check output quality on your own interactions. Speed alone does not guarantee good results.
-
Structured output: Email Topic Extraction and Email Generation must return JSON. Use a model verified for that prompt type. Models that are not verified may add extra text or formatting, or may not return JSON at all.
-
Custom redaction instructions: Instructions to exclude or redact information are best-effort with any model. They are not a guaranteed filter. In NiCE CXone testing, clearly structured instructions reduced leaked information more than switching models did. Structured instructions use separate lists of what to remove and what to keep, with examples. Review the output for sensitive information after any change to the prompt or the model.
Before You Publish a Profile That Uses a Different Model:
-
Test the prompt with a representative set of transcripts or email threads, including your longest and most complex ones.
-
Compare the output with NiCE Default output for the same interactions. Check the language, the required format, the content that must be included or excluded, the length, and the response time.
-
Switch only when the other model gives clearly better results for your use case.
-
Test again when you change the prompt, or when NiCE CXone announces that the model you selected will be retired.
If you select a model that is not verified for your prompt type, testing it for your use case is your responsibility.
Complete each of these tasks in the order given.
Launch GenAI Prompts
-
Click the app selector
. -
Under General, select AI Studio.
-
In the AI Studio left navigation, select GenAI Prompts.
Create Prompt Profile
Required permissions: GenAI Prompts, Prompt Access Edit
You will also need a license for NiCE CXone Copilot for Agents and the KB Query Extraction feature enabled.
Each flow comes with one default profile. You can create a custom prompt profile either by using a blank form or by copying an existing prompt and making edits. As you're working in the profile, it automatically saves for you only. Others cannot see the new profile, or edits you're making to an existing profile, until you click Publish Prompt. Profiles are also unavailable to select as the live prompt until they have been published.
If you create a new profile but don't publish it within 30 days of the most recent edit, the profile settings on the Prompt Configuration tab revert to their defaults. If you edit an existing profile but don't publish within 30 days of the most recent edit, changes on the Prompt Configuration tab revert to the settings specified in the latest published version.
-
In the Profiles tree expand one of these, depending on the interactions the prompt is for:
-
Copilot for Agents > KB Query Extraction.
-
Voice interactions: Copilot for Agents > KB Query Extraction (Voice).
-
Digital interactions: Copilot for Agents > KB Query Extraction (Digital).
-
-
Create a new profile. You can do this in one of two ways:
-
Create a new, blank profile. To do so, click the plus icon
next to the KB Query Extraction (Voice) or KB Query Extraction (Digital) folder. The new profile appears at the bottom of the list. -
Duplicate an existing profile.
-
Click the stacked dots icon
next to the existing profile you want to copy and select Duplicate. -
Rename the copy by locating it at the bottom of the profiles in that folder, clicking the stacked dots icon
, and selecting Rename. Enter the new name for the profile and click the check mark
to save it.
-
-
-
On the Prompt Configuration tab, adjust the values for these fields as needed:
-
What is the reasoning language?
-
What is the query language?
-
What is the tenant terminology and context?
-
What knowledge base scope?
Learn more about fields in this step
Field Details Reasoning Language Specifies the language used by the model to analyze the conversation and identify intent and key concepts. This setting affects internal reasoning only and does not control the language of the retrieved articles or generated responses.
It is recommended to choose the primary language used in customer or agent conversations.
Query Language Defines the language in which extracted queries are formulated before being submitted to the Knowledge Base. This setting determines which articles are most relevant during retrieval.
It is recommended to align this language with the language used to author Knowledge Base content.
Tenant terminology and context, also called tenant details Supplies tenant-specific language, acronyms, and contextual information that the model uses when forming Knowledge Base queries. Tenant details help normalize domain terms and reduce ambiguity during query extraction.
It is recommended to use product names, feature names, internal acronyms, abbreviations, brand-specific, or industry-specific terminology.
Knowledge base scope Controls which Knowledge Base sources are searched when executing extracted queries. Restricting the scope improves relevance by limiting results to applicable content.
It is recommended to select only the Knowledge Bases required for this prompt’s use case.
Broader scopes increase coverage but may reduce relevance.
-
-
Click Publish Prompt.
Add Transcripts for Testing
Required permissions: Launch GenAI Prompts, Prompt Access Edit
Import a Transcript
-
Click the Transcripts tab.
-
Under the Transcript Options area, in the Type of Transcript File field, select Segment.
-
Enter the segment ID of the transcript you want to use in the What are the segment IDs for the transcripts? text box. Click the add
button. The new transcript appears in the table below the text box. -
Click the paper and pencil icon
to view and edit the imported transcript. The transcript shows dialog from the agent in purple and the contact in blue. -
Modify the transcript as needed. You can take any of the following actions:
-
Click + Agent Phrase or + Client Phrase at the top of the window to add a new dialog box of a certain type. The new agent or contact dialog box appears at the bottom of the transcript. Enter the dialog for the role you selected.
-
Add another dialog box of the same type (agent or client) by selecting the box above where you want the new box and clicking the add dialog box icon
. The new box appears below the box you selected. Enter the dialog you want and then click the save icon
. -
Edit a dialog box by selecting the box you want to change and clicking the pencil icon
. Make the changes you need and then click the save icon
. -
Move dialog boxes higher or lower in the transcript by selecting the box and clicking the up icon
or the down icon
until the box is where you want it. -
Delete a dialog box by selecting the box and clicking the trash can icon
. -
Remove all dialog by clicking Clear in the top right corner of the window.
-
-
Continue to add agent and contact dialog until you have completed the transcript.
-
Click the Back to Transcript List icon
to return to the table of transcripts. -
Continue adding the transcripts you want to test. Testing the prompt will use all of the transcripts you have included in the table.
Manually Enter a Transcript
-
Click the Transcripts tab.
-
For the Type of Transcript File field, select New Transcript.
-
Click Create New Transcript. A new transcript appears in the table below.
-
Click the paper and pencil icon
. -
Modify the new transcript by using the following actions as needed:
-
Click + Agent Phrase or + Client Phrase at the top of the window to add a new dialog box of a certain type. The new agent or contact dialog box appears at the bottom of the transcript. Enter the dialog for the role you selected.
-
Add another dialog box of the same type (agent or client) by selecting the box above where you want the new box and clicking the add dialog box icon
. The new box appears below the box you selected. Enter the dialog you want and then click the save icon
. -
Edit a dialog box by selecting the box you want to change and clicking the pencil icon
. Make the changes you need and then click the save icon
. -
Move dialog boxes higher or lower in the transcript by selecting the box and clicking the up icon
or the down icon
until the box is where you want it. -
Delete a dialog box by selecting the box and clicking the trash can icon
. -
Remove all dialog by clicking Clear in the top right corner of the window.
-
-
Continue to add agent and contact dialog until you have completed the transcript.
-
Click the Back to Transcript List icon
to return to the table of transcripts. -
Continue adding the transcripts you want to test. Testing the prompt will use all of the transcripts you have included in the table.
Test and Publish the Prompt
Required permissions: Launch GenAI Prompts, Prompt Access Edit
After you have configured the prompt settings and added a transcript, you can test your prompt with that transcript. This will help you see what the AI-generated text looks like and gives you a chance to refine your settings. Publishing the prompt makes it available both for others to see and for selection in Copilot for Agents profiles.
-
When you are happy with your transcript, click Test Response.
-
Review the response generated for the customized prompt and transcript you provided. If you are not satisfied with the result, adjust the prompt configuration settings and try again. Continue testing and refining your settings until the extracted queries return the expected Knowledge Base results.
-
Click Publish Prompt.
Select the Prompt in Copilot for Agents
Once you have published the prompt profile, you can select it in the Copilot for Agents Config page profile.
- Click the app selector
and select General. - From the left menu, navigate to AI Studio > Profiles.
- From the list of Copilot profiles on the left, select the Copilot for Agents profile you want to apply the KB Query Extraction prompt profile to. If you don't have one yet, click Add Copilot Profile and follow these instructions to create one.
- On the Knowledge tab, scroll down to the CXone Knowledge Settings section and elect the prompt profile you want to use:
For voice interactions, from the KB Query Extraction (Voice) Prompt drop-down, select the prompt profile. This drop-down appears only when the voice channel is selected on the Configuration tab.
For digital interactions, from the KB Query Extraction (Digital) Prompt drop-down, select the prompt profile. This drop-down appears only when the digital channel is selected on the Configuration tab. The default is <Legacy>, which sends the contact's whole message to the knowledge base as the search query, without extracting a question.
- Click Save.