> ## Documentation Index
> Fetch the complete documentation index at: https://docs.caplena.com/llms.txt
> Use this file to discover all available pages before exploring further.
# How to Ensure High-Quality Results with the LLM
> Learn how to guide Caplena's LLM-based AI for precise topic assignments in just a few runs.
Caplena's **LLM-based AI** assigns topics faster, smarter, and with human-level accuracy. You no longer need endless manual reviews — what matters most now is how you **set up and refine your topics and descriptions**.
Here are our best tips to make sure your AI delivers excellent, trustworthy results.
## Tip 1: Understand how the LLM-based AI works
* The AI uses all available **project context** — project name, description, column titles, and of course, your text data.
* When you click **Done** in the Topic Assistant, the AI automatically:
1. Generates **topic descriptions** for each topic.
2. Assigns topics to your data using those descriptions.
3. Calculates an **AI Score** to indicate overall quality.
### Understanding Topic Assignment Consistency
The AI score reflects both the **quality** and **stability** of topic assignments. To produce reliable results, the model draws on your project background, examples, and topic descriptions, and runs the assignment multiple times — comparing results to arrive at the most consistent answer.
A **high score** means topics are well-differentiated, resulting in clear and stable decisions. A **lower score** suggests the boundaries between some topics may not be entirely clear-cut and could benefit from more distinct descriptions.
## Tip 2: Focus on Topics and their Descriptions
### Review & Adjust Your Topics
* Check for **duplicates or overlaps** — merge similar topics like *"Late Delivery"* and *"Shipping Delays"* into *"Delivery Issues".*
* Remove irrelevant or overly specific topics that don't add value.
* Add **missing topics** if you notice recurring ideas not yet covered.
Do all your topic cleanup **before** finalizing the topic collection. You'll save a rerun and improve the AI's learning context dramatically.
### Perfect Your Topic Descriptions
Topic descriptions are the **core input for the LLM-based AI**. The model reads these to understand what each topic means and uses them to decide how to assign responses.
Here's how to handle them effectively:
* Review all generated descriptions after the first run.
* Edit vague or generic ones, or delete confusing descriptions — the AI will automatically regenerate better ones during the next run.
* Avoid overlaps — make sure descriptions clearly distinguish similar topics.
## Tip 3: Do Smart Quality Checks
You no longer need to review hundreds of rows. Instead, use **targeted checks**:
1. **Check coverage** — if \~90% of rows have topics, you're in great shape.
2. **Spot-check 10–15 responses** — make sure assignments make sense.
3. **Use AI certainty** (optional) — rows with low confidence might reveal topics needing better descriptions.
4. **Focus on weak spots** — review topics with low coverage or unclear descriptions.
## Tip 4: Understand the AI Score
After each run you see an **AI Score**: how confidently and consistently the model assigned topics overall. While the theoretical maximum is 100, this is never achieved in practice — even skilled humans coding the same data twice don't fully agree with themselves.
Here's how to interpret it:
| Scenario | Typical score |
| :- | :- |
| Same person codes twice | \~90 |
| Two people code independently | 70–80 |
| Human-level AI performance | 70–85 |
The AI Score is **not** a percentage of responses assigned correctly — there's no external "right answer" it's checked against. A high score is a strong signal of quality, but it's not a substitute for the spot checks in Tip 3.
### If your score is lower than expected
The fastest path to a better score is usually **Tip 2** above — clearer, more distinct topic descriptions. A few extra things to check:
* **Overlapping topics** confuse the model just as much as they'd confuse a human reviewer. If two topics could plausibly apply to the same response, merge them or sharpen the distinction in their descriptions.
* **Low-confidence topics** are where the AI itself is least sure. Use the certainty view (Tip 3) to find them — they're the best candidates for a rewritten description rather than a blind rerun.
* **Sparse topics** (only a handful of responses) can show a missing or unreliable score simply because there isn't enough data for the model to gauge its own consistency. This isn't a quality problem to fix — it resolves itself as more matching data comes in, or you can merge the topic into a related one if it's not pulling its weight.
Rather not do this diagnosis by hand? Click the magic-wand icon () next to the AI Score to open the [Topic Critique Agent](/insight-and-reports/insight-agent/topic-critique-agent) — it runs this same analysis for you and proposes fixes. See the shortcut below for details.
## Tip 5: Refine Before Rerunning
Before running again, ask yourself:
* Are my topic descriptions clear and specific?
* Have I merged similar or overlapping topics?
* Have I removed or rewritten any confusing descriptions?
**Caplena supports both partial and full AI updates**:
* For small changes (e.g., updating 1–2 topics), a **partial update** is more efficient.
* For broader changes, a **full update** ensures consistency across your dataset.
Once you trigger a rerun, the AI will regenerate missing or deleted descriptions and reassign topics based on your updated setup.
## Shortcut: let Agent critique your topics for you
Tips 2–5 above describe the manual review process. If you'd rather have AI do the first pass, click the magic-wand icon () next to the **AI Score** to open the [Topic Critique Agent](/insight-and-reports/insight-agent/topic-critique-agent). It reviews your whole topic collection, explains what's hurting the score, and can apply description or structure fixes directly — with your confirmation before anything changes.
This is a different tool from the [Topic Assistant](/topic-assignment/understanding-the-basics/using-the-topic-assistant) panel in the Topics editor: the Topic Assistant surfaces new, similar, and rare topics as you build your structure, while the Topic Critique Agent specifically diagnoses and fixes what's dragging down your AI Score.
## Tip 6: Know When to Stop
Your project is ready when:
* The **AI Score** is 70+
* Each topic has a clear, correct description
* Most responses have a topic
* Spot checks look accurate
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.