> For the complete documentation index, see [llms.txt](https://docs.borndigital.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.borndigital.ai/insights/workspace/design/prompts.md).

# Prompts

This module is for creating and managing Prompts. They pair parameters into wholes that are then used for processing audio (transcriptions) and text inputs.

A prompt groups related [Parameters](https://docs.borndigital.ai/insights/workspace/design/parameters.md) into one request to the language model, for example all call-quality checks or all summary fields. Grouping saves cost and time because the model reads each record once per prompt, not once per parameter. Prompts also control the order of the analysis, when a step runs at all (conditions), and which model answers it.

### Project prompt

The collapsible **Project prompt** section at the top of the **Prompts** page holds LLM instructions for every prompt in the project, for example who the customers are and how to answer. A prompt without its own LLM instructions uses the project prompt; a prompt with its own instructions keeps them. Write the text in **Definition** and click **Save**. Each prompt's detail shows the project prompt read-only, with **Edit in Prompts**. Export and import of prompts include it.

<figure><img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2FmDrJXZqTrmmrQevuKyEN%2Finsights-prompts-project-prompt.png?alt=media" alt="The Prompts page of an Insights project with the expanded Project prompt section and its Definition field above the list of prompts"><figcaption><p>The Project prompt section on the Prompts page</p></figcaption></figure>

### Configuration

<img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2F9J29rm3prV3355BrLvif%2Fprompts-essential.png?alt=media" alt="Prompt — Essential configuration" width="375">

On-screen elements:

* **Essential**:
  * **Details**:
    * **Name**: input field of the prompt. The saved name is automatically normalized to lowercase `snake_case`.
    * **Language**: dropdown field for processing the prompt
    * **Prompt state**: radio buttons for choosing **Enabled**, **Disabled for all** (completely disabled), or **Disabled for rerun** (disabled only for reruns)
    * **Multi-chunk behavior**: radio buttons for choosing which parts of a long, multi-chunk record the prompt processes:
      * **Each** (default): processes every chunk. Use it when the analysis should cover the complete input, for example sentiment or topics across the whole conversation or document. The partial outputs are consolidated into one result according to the selected parameters' **Multi prompt behaviour**.
      * **First**: processes only the opening chunk. Use it for information expected near the beginning, for example a greeting, introduction, or initial reason for contact.
      * **Last**: processes only the final chunk. Use it for information expected near the end, for example a farewell, closing statement, or final outcome.

{% hint style="info" %}

#### Why the Multi-chunk behavior setting exists

Language models have limited context windows, so very long audio or text inputs may need to be broken into several chunks before processing. This prevents the end of a large document or recording transcript from being truncated.

Chunking is applied when the processed input exceeds 110,000 characters. There is no fixed recording-duration or word-count threshold:

* For **Audio projects**, it depends on the length of the generated transcript and its formatting, not directly on the recording duration.
* For **Text/Email projects**, it depends on the length and formatting of the processed text.

When an input contains multiple chunks, **Multi-chunk behavior** lets you decide whether the prompt processes every chunk, only the opening chunk, or only the final chunk. When **Each** is selected, the partial outputs are consolidated into a unified result according to the parameters' **Multi prompt behaviour**.
{% endhint %}

* **Parameters**:
  * **Predefined parameter group**: input field for selecting already predefined hardcoded parameter groups. Once any of them is selected, its predefined, hardcoded parameters are then listed in the **Parameters** field below.
  * **Parameters**: input field for selection of parameters (custom-made or predefined)

{% hint style="info" %}
Predefined groups and parameters come from the deployed configuration and are listed directly in the interface. Their availability can vary between environments and versions.
{% endhint %}

<details>

<summary>Predefined (hardcoded) parameter examples and their descriptions</summary>

The entries below are examples from a previously documented configuration. Available predefined parameters can vary between environments and versions; use the parameters listed in the current interface as the authoritative source.

<table data-header-hidden><thead><tr><th valign="top">Parameter</th><th valign="top">Definition</th><th valign="top">Constraint</th><th valign="top">Values</th><th valign="top">Output type</th><th valign="top">Fallback value</th><th valign="top">Multi Prompt Behaviour</th></tr></thead><tbody><tr><td valign="top">addressing</td><td valign="top">Value YES if the Agent used the Client’s first name or last name in the conversation. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">benefit</td><td valign="top">Did the agent mention the product benefit?</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">N/A</td><td valign="top">FIRST</td></tr><tr><td valign="top">call_script_summarization</td><td valign="top">Agent summarized Client's request to check if he understood correctly</td><td valign="top"><br></td><td valign="top">NO, YES</td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">call_script_telco_offering_accepted</td><td valign="top">Client accepted mobile phone service, package or tariff that Agent offered to him</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">call_script_telco_offering_value</td><td valign="top">Name of the offering, service, package or tariff that Agent offered to Client to buy, add or activate</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">LAST</td></tr><tr><td valign="top">category</td><td valign="top">Most dominant topic Client wants to discuss</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FREQUENT</td></tr><tr><td valign="top">category_product</td><td valign="top">categorize a product of selling</td><td valign="top"><br></td><td valign="top">insurance, savings, loans, investment, credit card</td><td valign="top">string</td><td valign="top">other</td><td valign="top">FIRST</td></tr><tr><td valign="top">client_behaviour_goodbye</td><td valign="top">Client's goodbye</td><td valign="top">Value YES if client said goodbye, else value NO</td><td valign="top">NO, YES</td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">LAST</td></tr><tr><td valign="top">client_behaviour_thanks</td><td valign="top">Client's client_behaviour_thanks</td><td valign="top">Value YES if client thanked for help, else value NO</td><td valign="top">NO, YES</td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">LAST</td></tr><tr><td valign="top">cross_sell</td><td valign="top">Value YES, if the Agent tried to sell more than just 1 product, or tried to cross sell, upsell another product. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">FIRST</td></tr><tr><td valign="top">empathetic</td><td valign="top">Value YES if the Advisor responds empathetically to the Clients concerns and questions. Value NO otherwise.</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top"><br></td><td valign="top">FIRST</td></tr><tr><td valign="top">extract_client_customer_contract_number</td><td valign="top">Client's customer number or customer contract number</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">extract_client_email</td><td valign="top">Client's email</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">extract_client_first_name</td><td valign="top">Client's first name</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">extract_client_insurance_number</td><td valign="top">Client's insurance number</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">extract_client_last_name</td><td valign="top">Client's surname</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">extract_client_phone_number</td><td valign="top">Client's phone number</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">extract_client_service_number</td><td valign="top">Client's service number</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">farewell</td><td valign="top">Value YES if the Agent said goodbye or wished the Client a nice day. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">future_interest</td><td valign="top">Value YES if the Client has expressed interest in a certain banking product in a certain time frame, such as another day, later, another time, sometimes in a few weeks, months, in half a year etc. Value NO if the Client has no interest in any banking product at all and when client does not mention interest then value No interest mentioned</td><td valign="top"><br></td><td valign="top">another day, later, another time, sometimes in a few weeks, months, in half a year, No interest</td><td valign="top">string</td><td valign="top">No interest mentioned</td><td valign="top">FIRST</td></tr><tr><td valign="top">future_interest_ii</td><td valign="top">Value YES if the Client has expressed interest in a certain banking product in a certain time frame, such as another day, later, another time, sometimes in a few weeks, months, in half a year etc. Value NO if the Client has no interest in any banking product at all and when client does not mention interest then value No interest mentioned</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">FIRST</td></tr><tr><td valign="top">gathered_category</td><td valign="top">Choose category that fit this conversation the most</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">greeted</td><td valign="top">The value YES, if the agent greeted, introduced himself, said the name of the company. The value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">FIRST</td></tr><tr><td valign="top">greeted_end_of_the_call</td><td valign="top">Yes , if agent greeted at the end of the call, otherwise NO</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">FIRST</td></tr><tr><td valign="top">handle_it</td><td valign="top">How well did the agent handle these objections? categorize it</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top"><br></td><td valign="top">FIRST</td></tr><tr><td valign="top">handle_objections</td><td valign="top">Value YES if the Agent is able to handle Customers objections well. NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">FIRST</td></tr><tr><td valign="top">identification</td><td valign="top">Value YES if the Client’s first name, last name and date of birth were mentioned in the conversation. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">interest</td><td valign="top">rate the clients interest about product and categorize it</td><td valign="top"><br></td><td valign="top">High, Moderate, Low</td><td valign="top">string</td><td valign="top">No interest</td><td valign="top">FIRST</td></tr><tr><td valign="top">interested_when</td><td valign="top">When will the Client be interested in the selected banking product? Return exactly one of the following options that is closest to what the Client said: Days, Week, 2 weeks, A few weeks, 3 weeks, Month, A few months, 2 months, 6 months, Year.</td><td valign="top"><br></td><td valign="top">Days, Week, 2 weeks, A few weeks, 3 weeks, Month, A few months, 2 months, 6 months, Year, Tomorrow</td><td valign="top">string</td><td valign="top">Not mentioned</td><td valign="top">FIRST</td></tr><tr><td valign="top">introduction</td><td valign="top">Value YES if the Agent introduced themselves with their first name, last name, and company name. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">FIRST</td></tr><tr><td valign="top">mifid_financial_products</td><td valign="top">Financial products that where mentioned or discussed in this Conversation by Agent of Client</td><td valign="top">string with values separated by comma</td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">CONCATENATE</td></tr><tr><td valign="top">mifid_financial_products_explained</td><td valign="top">If Agent explained risks or details about discussed financial products</td><td valign="top"><br></td><td valign="top">False, True</td><td valign="top">string</td><td valign="top">FALSE</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">mifid_relevant</td><td valign="top">True if conversation included any information that needs to be Mifid II regulated</td><td valign="top"><br></td><td valign="top">False, True</td><td valign="top">string</td><td valign="top">FALSE</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">obscene</td><td valign="top">All profanity, obscene words that are in the transcript. NOT_FOUND if none occur.</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top"><br></td><td valign="top">FIRST</td></tr><tr><td valign="top">other_help</td><td valign="top">Value YES if the Agent asked the Client whether they had any other requests or if there was any other way they could help. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">personalizing_offer</td><td valign="top">Value YES if the Agent tried to personalizing the offer to Clients needs/situation, rather then just presenting product as on a leaflet. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">NO, YES</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">FIRST</td></tr><tr><td valign="top">react_to_offer</td><td valign="top">How did the customer react to the offer? categorize it</td><td valign="top"><br></td><td valign="top">Positive, Interest, Disinterested, Disappointed, Polite, Neutral</td><td valign="top">string</td><td valign="top">Not mentioned</td><td valign="top">FIRST</td></tr><tr><td valign="top">recapitulation</td><td valign="top">Value YES if the Agent summarized the conversation and verified the contact details with the customer. Value NO otherwise.</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">NO</td><td valign="top">WEIGHTED_VALUE</td></tr><tr><td valign="top">rejection</td><td valign="top">List the main reason why the client does not want to purchase the product or service, why they are not interested.</td><td valign="top"><br></td><td valign="top">Price Objection, Lack of Need, Competitor Preference, Budget Constraints, Lack of Understanding, Not the right time</td><td valign="top">string</td><td valign="top">Not mentioned</td><td valign="top">FIRST</td></tr><tr><td valign="top">rejection_reason_description</td><td valign="top">Summarize the main reason why the client does not want to purchase the product or service, why they are not interested.</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top"><br></td><td valign="top">FIRST</td></tr><tr><td valign="top">selling</td><td valign="top">Did the agents sell the product?</td><td valign="top"><br></td><td valign="top">YES, NO</td><td valign="top">string</td><td valign="top">N/A</td><td valign="top">FIRST</td></tr><tr><td valign="top">sentiment_agent</td><td valign="top">Agent's sentiment at start of conversation</td><td valign="top"><br></td><td valign="top">Helpful, Friendly, Unfriendly, Aggressive, Neutral, Other</td><td valign="top">string</td><td valign="top">Neutral</td><td valign="top">FREQUENT</td></tr><tr><td valign="top">sentiment_client_end</td><td valign="top">Client's sentiment at end of conversation</td><td valign="top"><br></td><td valign="top">Relief, Gratitude, Frustration, Confusion, Anger, Neutral, Other</td><td valign="top">string</td><td valign="top">Neutral</td><td valign="top">LAST</td></tr><tr><td valign="top">sentiment_client_start</td><td valign="top">Client's sentiment at start of conversation</td><td valign="top"><br></td><td valign="top">Relief, Gratitude, Frustration, Confusion, Anger, Neutral, Other</td><td valign="top">string</td><td valign="top">Neutral</td><td valign="top">FIRST</td></tr><tr><td valign="top">subcategory</td><td valign="top">Subcategory to the most dominant topic Client wants to discuss</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FREQUENT</td></tr><tr><td valign="top">subsubcategory</td><td valign="top">Subcategory to the most dominant topic Client wants to discuss</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FREQUENT</td></tr><tr><td valign="top">suggest_category</td><td valign="top">Most dominant topic Client wants to discuss</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FREQUENT</td></tr><tr><td valign="top">suggest_subcategory</td><td valign="top">Subcategory to the most dominant topic Client wants to discuss</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FREQUENT</td></tr><tr><td valign="top">summary</td><td valign="top">Summary of what was Conversation about</td><td valign="top">Maximum 20 words</td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">CONCATENATE</td></tr><tr><td valign="top">summary_long</td><td valign="top">Summary of what was Conversation about</td><td valign="top">Maximum 40 words</td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">CONCATENATE</td></tr><tr><td valign="top">summary_short</td><td valign="top">Summary of what was Conversation about</td><td valign="top">Maximum 10 words</td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">CONCATENATE</td></tr><tr><td valign="top">topic_dominant</td><td valign="top">Most dominant topic this Conversation was about</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FREQUENT</td></tr><tr><td valign="top">topic_first</td><td valign="top">First topic Client wants to discuss</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FIRST</td></tr><tr><td valign="top">topic_nondominant</td><td valign="top">Second most dominant topic this Conversation was about</td><td valign="top"><br></td><td valign="top"><br></td><td valign="top">string</td><td valign="top">NOT_FOUND</td><td valign="top">FREQUENT</td></tr></tbody></table>

</details>

* **Previous parameter result**: input field on the **Essential** tab for selecting a result from an earlier prompt and using it as input for the current prompt.\
  *Example: summary of the audio record (`prompt_summary`)*

{% hint style="info" %}
You can select parameter results from prompts that run earlier. The available name uses the format `<prompt>_<parameter>`.
{% endhint %}

<img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2FllD91POyqhuqLv6cUvmZ%2Fprompts-advanced.png?alt=media" alt="Prompt — Advanced configuration" width="375">

* **Advanced**:
  * **Analysis enhancement**:
    * **LLM instructions**: input field for additional prompt specifics. E.g., output file type, processing specifics, structure requirement, etc. It is mainly used in some specific cases to fine-tune the output. When it is empty, the prompt uses the [project prompt](https://docs.borndigital.ai/insights/workspace/design/prompts.md#project-prompt).
  * **Conditions**: use conditions to run a prompt only when selected values match the configured pattern. A prompt without conditions always runs.
    * **Condition**: dropdown that chooses how multiple conditions are combined.
      * **AND** (default): all conditions must match.
      * **OR**: at least one condition must match.
    * **Add new condition**: adds another condition. The Bin icon next to a condition removes it.
      * **Parameter**: the exact, case-sensitive name of a value available to the prompt. The field suggests the `<prompt>_<parameter>` names of parameters from earlier prompts; you can also type another name.
        * Analytical parameter results from earlier prompts use the available `<prompt>_<parameter>` name.
        * Metadata values normally use an available `meta_input_<metadata_name>` name.
        * A parameter name is not a regex or a path into nested metadata. A dot does not select a value inside a nested object.
          * Example: `meta_input_order_details`**`.channel`**
      * **Value**: enter basic text or a regular expression (regex) to match the selected parameter. Enter a regex directly, without `/` characters around it. When the selected parameter has a **List of values**, the field suggests those values.
        * **Basic text matching (default):** `phone` matches any existing value that contains `phone`, such as `phone` or `smartphone`.
        * **Exact matching:** `^phone$` matches only when the complete value is `phone`.
        * **Case-insensitive exact matching:** `(?i)^phone$` matches `phone`, `Phone` or `PHONE`.
        * **Example — use nested metadata:**
          * Suppose the source metadata contains:

            ```json
            {
              "order_details": {
                "channel": "phone"
              }
            }
            ```
          * The desired nested field, `meta_input_order_details.channel`, is not available as a separate **Parameter** unless the source metadata was already flattened into a field with that exact name.
          1. In **Parameter**, enter `meta_input_order_details`.
          2. In **Value**, use one of these patterns:
             * `phone` to match this value anywhere in the nested metadata.
             * `['"]channel['"]\s*:\s*['"]phone['"]` to match the specific `channel` field with the value `phone`.
          3. The prompt runs only when the selected pattern matches.
        * **Important:**
          * A blank **Parameter** makes that condition pass. With **OR**, one blank condition can make the complete condition group pass.
          * A blank **Value** matches every existing value. A missing parameter or a parameter with no value does not match.
          * Patterns use RE2 regular expressions. They are not wildcards: write `.*` where you would write `*`. Lookarounds and backreferences are not supported.
          * Escape regex characters such as `.`, `*`, `+`, `?`, `(`, `)` and `[` when they should be treated as ordinary characters.
          * A prompt with an invalid pattern cannot be saved; the error message names the condition's parameter.
          * Matching nested metadata is less reliable than using a flattened field. Where possible, use a dedicated field such as `meta_input_channel`.
  * **Custom provider**: switch for using a different LLM provider or model for this prompt. When it is off, the prompt uses the project's or organization's default provider, which Born Digital sets up. When you switch it on, **Provider** is preset to `azure`; fill in the settings below.
    * **Provider**: `anthropic`, `azure` (Azure OpenAI), `gemini`, `groq` or `openai`
    * **API Key**: input field for the provider's key. A stored key is shown as `*****` and is never sent back to the browser. Leave the field untouched to keep the stored key, or click into it and type a new one to replace it.
    * **Model name**: input field
    * **API base**: input field for entering a URL
    * **API version**: input field
    * **Reasoning effort**: dropdown with **none**, **minimal**, **low**, **medium**, **high**, and **xhigh**. Which values a model accepts depends on the model.
  * **LLM caching**: section that decides what the LLM provider caches. Prompt caching reuses the part of a request the provider has already processed, which lowers cost and speeds up answers when many requests start the same way. It has two dropdowns:
    * **LLM caching**:
      * **Project default** (default): uses the caching setting of the project.
      * **Conversation first**: caches the record (the conversation or text), so all prompts that run over the same record reuse it. Suitable when a record goes through many prompts.
      * **Prompt first**: caches this prompt, so it is reused for all records. It also enforces strict JSON output.
    * **Explicit prompt cache**: **Project default** (default), **On** or **Off**. **On** marks exactly which part of the request the provider should cache instead of leaving it to the provider. As the helper text under the field says, only newer models support it and older ones reject the request, so turn it on only after checking that your model supports it.

<figure><img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2Fgit-blob-5a82faa3343d078c2297d724343bb972f7ac16b1%2Finsights-prompt-caching.png?alt=media" alt="The Advanced tab of a prompt with Analysis enhancement, Conditions, Custom provider and the open LLM caching list Project default, Conversation first and Prompt first highlighted"><figcaption><p>LLM caching on the Advanced tab</p></figcaption></figure>

* **Create** / **Save**: button at the bottom for creating the prompt or saving changes to it. **Name** and at least one **Parameter** are required.
* **Delete**: button at the bottom for deleting the prompt. It is active only for a saved prompt.

{% hint style="info" %}
Tip: Use the predefined parameters for inspiration, but add little details for your individual use-case. (i.e., the exact call script for your agents to introduce to customers, like: The value YES, if the agent greeted, introduced himself and asked "How can I help you?". The value NO otherwise.

vs. the default:

The value YES, if the agent greeted, introduced himself, said the name of the company. The value NO otherwise.
{% endhint %}

### Management

* The order of prompts defines their running order. Drag prompts to reorder them, then select **Save**. Deleting prompts and changing their state in the list also take effect only after **Save**.

<figure><img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2FRQWqrXCV3gG3c3ahGa4R%2Fimage.png?alt=media&amp;token=3d0dec66-6393-4a55-b3df-4380eaa00a4e" alt=""><figcaption></figcaption></figure>

On-screen elements:

* Upper part:
  * **Search prompts**: bar for searching prompts by name
  * **+ Create**: button for creating a new prompt
  * **Project prompt**: collapsible section with the [project prompt](https://docs.borndigital.ai/insights/workspace/design/prompts.md#project-prompt)
  * **Select**: enters bulk selection mode
    * **Select all**: selects every prompt in the current list
    * **Done**: exits bulk selection mode
    * **Prompts state**: dropdown field for changing the state of the selected prompts to one of these: **Enabled**, **Disabled for all** (completely disabled), or **Disabled for rerun** (disabled only for reruns)
    * **Delete**: button for deleting the selected prompt(s)
* List of prompts: *(Elements are listed from the left)*
  * Drag handle: for changing the running order
  * **Name**: of the prompt
  * **Pencil**: icon button for editing the prompt
  * State icon: shows the prompt state; hover over it to see the state's name
  * **Language**: used for processing the prompt
  * **Prompt provider type**: displaying **Default provider** or the name of the custom provider
  * **Parameters preview**: the parameters used by the prompt. Hover over one to see its definition, or click it to open the parameter.
  * Bin icon: deletes the prompt
* Floating buttons:
  * **Export**: downloads a JSON file with the project's prompts, the parameters created in this project, and the project prompt. **Custom provider** settings are not exported, so API keys never leave the project.
  * **Import**: loads such a JSON file. It creates the parameters from the file in this project and **replaces the project's prompts** with the imported ones, including the project prompt when the file contains one. Imported prompts have no custom provider. If a parameter used by an imported prompt already exists with the same name, the import stops and lists the duplicated names.
  * **Save**: saves the changes made in the list: the order, prompt states, deletions, and the project prompt. Until you select **Save**, nothing in the list is changed.
