LLM Prompt¶

Description¶

The LLM Prompt Node lets you use prompts with different LLM models to generate text or structured content. This Node combines the capabilities of the AI Agent Node and the legacy LLM Prompt Node by adding support for tools, including MCP tools. The Node also offers more customization while avoiding built-in personalization of your AI Agent. You can personalize the prompt directly if needed. Use this Node together with the Load AI Agent Node, store your AI Agent configuration in the Input or Context object, and reuse it in the LLM Prompt Node.

Before using this Node, set the LLM provider in the Settings.

Parameters¶

Large Language Model

The selected Default model is the model that you specified in Settings > Generative AI Settings of your Project.

You can select a different model from the list or override the selected model using the Custom Model Options parameter.

System Prompt

The system prompt is the message sent to the LLM to guide its responses. The parameter supports CognigyScript, allowing dynamic content and logic within the prompt. This prompt can work either as the input for completion tasks or as the system message in chat-based interactions, setting context and behavior for the AI Agent.

Additionally, you can inject the recent conversation into the System Prompt field by using these tags:

@cognigyRecentConversation — the tag is replaced with a string that can contain up to 10 recent AI Agent and 10 user outputs, for example:
```
Bot: agentOutput1
User: userOutput1
Bot: agentOutput2
User: userOutput2
```
@cognigyRecentUserInputs — the tag is replaced with a string that can contain up to 10 recent user outputs, for example:

User: userOutput1
User: userOutput2

If you want to access only the last user input, specify the Text token in the System Prompt field.

When adding a tag, ensure that you leave a line break before and after the tag, for example:

A user had a conversation with a chatbot. The conversation history so far is:
@cognigyRecentConversation

Describe the user sentiment in one very short line.

Both tags can include an optional turn limit parameter, which is appended to the tag.

Examples:

@cognigyRecentConversation:3 // returns the last 3 turns of the conversation.
@cognigyRecentUserInputs:2 // returns the last 2 user inputs.

Advanced

Parameter	Type	Description
Maximal Tokens	Slider	The maximum number of tokens to generate in the completion.
Use Single Prompt Mode	Toggle	Send a single prompt to the model without any conversation context. This parameter is disabled by default. It doesn't support multi-turn conversations or chat and is useful for simple, one-off completions.
Transcript Turns	Slider	The number of conversation turns to include in the LLM chat completion request. By default, the value is `50`.
Response Format	Select	Choose the format for the model's output result. You can select one of the following options: None — no response format is specified, or do not request with an LLM provider that does not accept any response format or does not support it or could use provider's default in-built response format. This option is selected by default. Text — the model returns messages in text format. JSON Object — the model returns messages in JSON format. To use this option, instruct the model to generate JSON via a system or user message in the Instruction (System Message/Prompt) field. For example, `Take the user message and return it as JSON in the following format {"message": "THE_MESSAGE"}`. Note that this parameter may not be supported by all LLMs. For more information, refer to the LLM provider's API documentation.
Timeout	Number	The maximum number of milliseconds to wait for a response from the Generative AI Provider.
Sampling Method	Select	Methods: Temperature — determines the level of randomness in the generated text. A higher temperature allows for more diverse and creative outputs, while a lower temperature leads to more predictable and consistent outputs with the training data. Top Percentage — specifies the percentage of the most probable outputs for generation, resulting in more consistent output.
Temperature	Slider	Define the sampling temperature, which ranges between 0 and 1. Higher values, such as 0.8, make the output more random, while lower values, such as 0.2, make it more focused and deterministic.
Top Percentage	Slider	Control the Top-p (nucleus) sampling, ranging from 0 to 1. Higher values allow more diverse word choices, while lower values make the output more focused. For example, 0.9 means the model selects from the smallest set of words with a combined probability of 90%.
Presence Penalty	Slider	Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood of talking about new topics.
Frequency Penalty	Slider	Number between -2.0 and 2.0. The penalty assigns a lower probability to tokens frequently appearing in the generated text, encouraging the model to generate more diverse and unique content.
Use Stops	Toggle	Whether to use a list of stop words to let Generative AI know where the sentence stops.
Stops	Text	Up to 4 sequences where the API will stop generating further tokens. The returned text will not contain the stop sequence.
Seed	Number	Use this parameter for consistent output when referring to the same LLM Prompt Node multiple times. Specify any integer number, for example, `123`. The number in the Seed field and the prompt in the Instruction (System Message/Prompt) field should remain unchanged during subsequent references to the Node. Note that in OpenAI, this parameter is in Beta and is supported only by certain models.

Storage & Streaming Options

Parameter	Type	Description
How to handle the result	Select	Determine how to handle the prompt result: Store in Input — stores the result in the Input object. To print the prompt result, use the LLM Prompt Result Token in the Say Node or enable the Output result immediately option. Store in Context — stores the result in the Context object. To print the prompt result, use the LLM Prompt Result Token in the Say Node or enable the Output result immediately option. Stream to Output — streams the result directly into the output. This means that the model provides prompts directly into the conversation chat, as soon as a Stream Buffer Flush Token is matched, and you don't need to use the LLM Prompt Result token and Say Node. By default, this result won't be stored in either the Input or the Context. You can change this behavior by activating the Store Copy in Input option. Note that streaming may not be supported by all Cognigy LLM Prompt Node providers, such as Google¹. If streaming is not supported, enabling the Store Copy in Input option will save the result to the Input object.
Input Key to store Result	CognigyScript	The parameter appears when Store in Input is selected. The result is stored in the `promptResult` Input object by default. You can specify another key.
Context Key to store Result	CognigyScript	The parameter appears when Store in Context is selected. The result is stored in the `promptResult` Context object by default. You can specify another key.
Stream Buffer Flush Tokens	Text Array	The parameter appears when Stream to Output is selected. It defines tokens that trigger the stream buffer to flush to the output. The tokens can be punctuation marks or symbols, such as `\n`.
Stream Buffer Flush Overrides	Text Array	The parameter appears when Stream to Output is selected. It allows using regular expressions (without leading or trailing slashes) to control stream buffer flushing. A trailing `$` is automatically added to match patterns at the end of the buffer. For example, `\d+\.` checks for a number followed by a dot at the end of the string.
Output result immediately	Toggle	The parameter appears when you select either Store in Input or Store in Context. This parameter allows you to output results immediately without using the Say Node and LLM Prompt token.
Store Detailed Results	Toggle	The parameter appears when you select either Store in Input or Store in Context, or when you enable Store Copy in Input. This parameter allows you to save detailed results of the LLM's generated output. By default, the result is stored in the `promptResult` object. You can specify another value in the Context Key for the Result field to save it in the Context object, or in the Input Key for the Result to save it in the Input object. The object contains keys such as `result`, `finishReason`, and `usage`. It may also include `detailedResult` if completion models are used, as well as `firstChunk` and `lastChunk` in some streaming results, depending on the LLM provider.
Store Copy in Input	Toggle	The parameter appears when Stream to Output is selected. In addition to streaming the result to the output, store a copy in the Input object by specifying a value in the Input Key to store Result field.

Streaming Results

In Stream mode, LLM generates tokens and returns them one by one to Cognigy.AI to ensure low-latency responses. Cognigy.AI monitors delimiter tokens (Stream Buffer Flush Tokens), which serve as markers indicating when to output the token buffer. These tokens could be ., !, ?, or any other symbols that act as delimiters for complete logical statements. When Cognigy.AI detects one of these tokens, it promptly flushes the token buffer into the voice or text chat.

The preconfigured overrides are listed in the table.

Regex	Description	Example
`\d+\.`	A number followed by a dot.	26.08
`\b(?:Dr\|Ms\|Mr\|Mrs\|Prof\|Sr\|Jr\|ca)\.`	Common abbreviations followed by a dot.	Mr.
`\b[A-Z]\.`	A single capital letter followed by a dot.	M. Smith
`\.\.\.`	Three dots used for omission.	...
`\b.\..\.`	Two-letter abbreviations.	i.e., e.g.

Tool Settings

The process of setting up a tool is the same as for the AI Agent Node. See the example in the AI Agent Tool Settings section.

Parameter	Type	Description
Tool Choice	Selector	If supported by your LLM Model, this will determine how tools should be selected by the AI Agent: Auto — tools (or none) are automatically selected by the AI Agent when needed. Required — your AI Agent will always use one of its Tools. None — your AI Agent won't use a tool.
Use Strict mode	Toggle	When the parameter is enabled, strict mode (if supported by the LLM provider) ensures that the arguments passed to a tool call precisely match the expected parameters. Enabling this feature can help prevent errors. However, it may cause a slight delay in the response, especially during the first call after making changes.

Image Handling

Parameter	Type	Description
Process Images	Toggle	Enable the AI Agent to read and understand images attachments. Make sure that your LLM provider supports image processing; refer to your provider's documentation. In addition, make sure that attachments are supported by and activated in your Endpoint, for example, Webchat.
Images in Transcript	Selector	Configure how images older than the last turn are handled to reduce token usage: Minify — reduces the size of these images to 512x512px. Drop — excludes the images. Keep — sends the max size (this option consumes more tokens). Limitations and token consumption depend on the LLM used.

Error Handling

Parameter	Type	Description
Log to System Logs	Toggle	Log errors to the system logs. They can be viewed on the Logs page of your Project. The parameter is inactive by default.
Select Error Handling Approach	Select	You can select one of the Error Handling options: Stop Flow Execution — terminate the current Flow execution. Continue Flow Execution — allow the Flow to continue executing, bypassing the error and proceeding to the next steps. Go to Node — redirect the workflow to a specific Node in the Flow, which can be useful for error recovery or customized error handling.
Error Message (optional)	Text	The parameter appears when Continue Flow Execution is selected. Add an message to output if the LLM Prompt Node fails.
Select Flow	Select	The parameter appears when Go to Node is selected. Select a Flow from the available options.
Select Node	Select	The parameter appears when Go to Node is selected. Select a Node from the available options.

Custom Options

These settings are helpful if you need to use parameters that are not included in the LLM Prompt Node or if you need to overwrite existing ones.

Parameter	Type	Description
Custom Model Options	JSON	Additional parameters for the LLM model. You can specify individual parameters as well as entire functions. These parameters customize the behavior of the model, such as adjusting temperature, top_k, or presence_penalty. Note that if you use a parameter already set in the Node, for example, temperature, it will be overwritten. To view the full list of available parameters for your model, refer to the LLM provider's API documentation. For example, OpenAI or Azure OpenAI. Examples: `{ "logprobs": true }` `{ "temperature": 0.7 }` `{ "model": "claude-3-opus-20240229" }` (see Forcing Model Versions)
Custom Request Options	JSON	Additional parameters for the LLM request. These parameters customize the request itself, such as setting parameters related to timeout, retries, or headers. For more information, refer to the LLM provider's API documentation. Examples: - `{ "timeout": 5000 }` - `{ "headers": { "Authorization": "Bearer <token>" } }`

Forcing Model Versions

You can force the LLM Prompt Node to use a specific model version by including it in the Custom Model Options. This means that the LLM Prompt Node will use the specified version of the language model instead of the default or any other available versions. This allows for more control over the behavior of the LLM Prompt Node, ensuring it utilizes the desired model version for generating prompts or responses.

You can use models from any LLM provider supported by Cognigy, including those not yet directly integrated. However, you can only replace one model with another within the same provider.

Let's consider an example with the Anthropic provider: how to force the LLM Prompt Node to use the model version claude-3-sonnet-20240229, despite the LLM resource defaulting to the claude-3-opus-20240229 model:

Create an Anthropic LLM resource for Claude-1, for example, claude-3-opus-20240229.
Create a Flow and add an LLM Prompt Node to it.
In the LLM Prompt Node, select the model claude-3-opus-20240229 from the Large Language Model list.
Override the model selection. In the Custom Model Options field, specify the custom model options as follows: { "model": "claude-3-sonnet-20240229" }.
Click Save Node.

The LLM Prompt Node now utilizes the claude-3-sonnet-20240229 model.

Below, you'll find documentation for supported models:

Debugging Settings

When using the Interaction Panel, you can trigger two types of debug logs. These logs are only available when using the Interaction Panel and are not intended for production debugging. You can also combine both log types.

Parameter	Type	Description
Show Token Count	Toggle	Send a debug message containing the input, output, and total token count. The message appears in the Interaction Panel when Debug Mode is enabled. Cognigy.AI uses the GPT-3 tokenizer algorithm, so actual token usage may vary depending on the model used. The parameter is inactive by default.
Log Request and Completion	Toggle	Send a debug message containing the request sent to the LLM provider and the subsequent completion. The message appears in the Interaction Panel when Debug Mode is enabled. The parameter is inactive by default.
Log Tool Definitions	Toggle	Send a debug message containing information about the configured AI Agent tools. The message appears in the Interaction Panel when debug mode is enabled. The parameter is inactive by default.

More Information¶

Note that not all LLM models support streaming. ↩