# Chat Best Practices

{% hint style="info" %}
Effective prompting has a learning curve.  Crafting clear, concise prompts can significantly enhance the quality of the results you receive from Lumi.
{% endhint %}

Getting the most out of Lumi starts with crafting strong, clear prompts. The quality of your prompts directly affects the relevance, accuracy, and completeness of the insights you receive. This guide outlines five essential best practices, grounded in real-world usage and training sessions.

### Lumi Best Practices Overview

| Best Practice Category                                                        | Understanding the Rationale                                                                                      |
| ----------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------- |
| [Understand What Data is Available](#id-1.-understand-what-data-is-available) | You need to know both what data exists and how to ask for it effectively.                                        |
| [Use Clear and Specific Language](#id-2.-use-clear-and-specific-language)     | Specificity drives quality; vague requests often lead to irrelevant or incomplete results.                       |
| [If Required, Specify Columns](#id-3.-if-required-specify-columns)            | When applicable, naming the exact columns helps Lumi retrieve more accurate and relevant outputs.                |
| [Leverage Multi-Prompt Approach](#id-4.-leverage-a-multi-prompt-approach)     | Use a two-prompt strategy: first locate the values or identifiers, then use them in your second analysis prompt. |
| [Validate Assumptions & Approach](#id-5.-validate-assumptions-and-approach)   | Check explanation and assumptions to ensure accuracy.                                                            |

***

### 1. Understand What Data is Available

Before diving into complex analysis, begin by exploring your data. Lumi can help you preview tables, inspect distinct values, check ranges, and understand distributions.

**Why this matters:** Prompts grounded in data context produce better results. If you’re unsure about the data structure, you're more likely to get irrelevant or incomplete responses.

<figure><img src="https://1092914297-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FbsNNtXffkOLYrtYRQIcx%2Fuploads%2FPSGsp0ooz5CfXxbNXQK8%2FScreenshot%202025-06-03%20at%202.52.51%E2%80%AFPM.png?alt=media&#x26;token=d5b7f415-3e3b-4484-a320-554bb22c0997" alt=""><figcaption></figcaption></figure>

**Example prompts:**

* `Show me 100 random records from the sales header table.`
* `What are the distinct customer IDs in the customer_details table?`
* `What is the count of distinct product categories in the sales header table?`
* `What is the earliest and latest order date in the sales header table?`

***

### 2. Use Clear and Specific Language

Avoid shorthand or vague phrasing. Be as specific and descriptive as possible.

**Why this matters:** Specificity improves Lumi’s ability to interpret and return relevant answers. General prompts create room for incorrect assumptions.

<figure><img src="https://1092914297-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FbsNNtXffkOLYrtYRQIcx%2Fuploads%2FfVUhuTb33hbSzTa6OsVL%2FScreenshot%202025-06-03%20at%202.52.56%E2%80%AFPM.png?alt=media&#x26;token=734e6b71-91d5-4e02-8001-d743820b6b06" alt=""><figcaption></figcaption></figure>

**Example:**

* <mark style="color:red;">Vague prompt:</mark> `Top products 2023?`
* <mark style="color:green;">Specific prompt:</mark> `What are the top 25 products by revenue in 2023? Sort descending.`

***

### 3. If Required, Specify Columns

When requesting comparisons or multi-metric outputs, include the exact fields you want to see. This gives Lumi guidance to structure the query appropriately.

**Why this matters:** When Lumi knows what output format you want, it's more likely to return usable and correctly scoped results.

<figure><img src="https://1092914297-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FbsNNtXffkOLYrtYRQIcx%2Fuploads%2F1Q7kqYcg8FIxF8wFLAo8%2FScreenshot%202025-06-03%20at%202.53.07%E2%80%AFPM.png?alt=media&#x26;token=862f7815-9035-460e-853b-27dee981206c" alt=""><figcaption></figcaption></figure>

**Example:**

* <mark style="color:orange;">Unspecified prompt:</mark> `Top 10 brands by revenue growth in Jan 2024`
* <mark style="color:green;">Specific Column Prompt:</mark> `Top 10 brands by revenue growth in Jan 2024.`` `*`Output should include brand, January 2024 revenue, January 2023 revenue, and the delta.`*

{% hint style="warning" %}
**Tip:** You can also ask Lumi to include “relevant details” if you're unsure which supplemental columns might be helpful.
{% endhint %}

***

### 4. Leverage a Multi-Prompt Approach

Break more complex tasks into multiple steps. Use initial prompts to gather identifiers or filters, then follow up with specific analysis using those values.

**Why this matters:** Lumi does not retain prompt memory. Each question should include all necessary context.

<figure><img src="https://1092914297-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FbsNNtXffkOLYrtYRQIcx%2Fuploads%2FbiEBxD6fBFaG9NUm6oOH%2FScreenshot%202025-06-03%20at%202.53.11%E2%80%AFPM.png?alt=media&#x26;token=a818de8b-c277-4845-b333-78710ab01c7b" alt=""><figcaption></figcaption></figure>

**Example prompt strategy:**

> * **Prompt 1:** List the top 5 customer segments by revenue in Q1 2025.
>
>   **Customer Segments Return:** Churned, <mark style="color:purple;">New</mark>, Returning, VIP, <mark style="color:green;">Loyal</mark>, <mark style="color:orange;">High Value.</mark>&#x20;
> * **Prompt 2:** Compare monthly transaction trends for segments: <mark style="color:green;">Loyal</mark>, <mark style="color:orange;">High Value</mark>, and <mark style="color:purple;">New</mark>.

***

### 5. Validate Assumptions and Approach

Review the explanations, logic, and SQL that Lumi provides. Ensure that the metrics, filters, and joins used align with your intended analysis.

**Why this matters:** Lumi may make assumptions, especially around calculations nuances (timelines, type of comparisons, filters etc). You can review to understand if this was the correct logic on the output.

<figure><img src="https://1092914297-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FbsNNtXffkOLYrtYRQIcx%2Fuploads%2Fy9KsCadruKRdJJF6ODcf%2FScreenshot%202025-06-03%20at%202.53.15%E2%80%AFPM.png?alt=media&#x26;token=e0cb3260-12f8-4128-9594-974f538c818d" alt=""><figcaption></figcaption></figure>

**Checklist:**

* Read the assumptions and explanation tab.
* Confirm filters and logic are appropriate.
* Use thumbs up/down to provide feedback and help refine future responses.

***

#### Known Limitations

Lumi is designed for structured, insight-driven analysis. However, there are several limitations to be aware of when crafting prompts see [chat limitations](https://docs.lumi-ai.com/using-lumi/best-practices/broken-reference) for more information:

* **Data Availability:** Lumi can only query data that is connected and defined in the Knowledge Base. If a field or table doesn’t exist in the model, Lumi won’t be able to retrieve insights; even if the prompt is well-written.
* **Data Returned From Query:** Lumi cannot see the actual table output from prior prompts. It only remembers the previous question and a summary. You must restate any specific values or identifiers in your follow-up questions.
* **Follow Up Questions:** While Lumi may generate follow-up suggestions, these are not always schema-aware. They can include filters or fields that don’t exist in your data model. Manual rephrasing is often necessary.
* **Row Limitations:** Lumi enforces a 100-row maximum per query to maintain performance and resource efficiency. Large result sets should be filtered, limited, or aggregated before requesting.
* **Latency:** Lumi’s query speed is directly influenced by your database's performance. Slow queries typically result from limited compute capacity.

***

### Summary

Prompting well is about clarity, specificity, and structure. Understand your data, ask precise questions, and always verify assumptions. With these practices, you’ll unlock more valuable and accurate insights using Lumi.

{% hint style="success" %}
**Clarity = Better Output**
{% endhint %}