How the new LLM Topic Assignment works

Here are some tips to help you improve the performance of the new AI system for topic assignment.

Project & Question Context

In addition to text + reviewed texts, the AI now uses the following fields during training:

Project Title
Project Description

Question Name
Question Description

Best Practices

Use descriptive titles instead of generic ones if available
Example: Use “Feedback on Delivery Process” instead of “Q3”.
Add contextual info to descriptions:
- Lesser-known companies or terms
  Example: “XXX is a subsidiary of YYY that focuses on manufacturing.”
- Technical or company-specific terms
  - Define acronyms or jargon that might not be widely known.
  - Generally, the AI will be able to understand technical terms even if they are only known to specialists in the field. However, if the terms are uncommon acronyms or are company-specific, it may be useful to define them.
Sentiment guidance
Example: If responses are from complaints, say so — this helps the AI interpret tone or sarcasm better.

Topic Descriptions

Topic descriptions are now a major input for the AI.

These are auto-generated based on context (topic name + project/question info)
But… they may be wrong or unclear if topic names are vague

When to Update Descriptions

A topic is frequently misassigned or missed
The topic name is unclear or ambiguous
Example: “This topic solely covers the quality of the food. It should NOT include comments about cost or value.”

Notes on Behavior

Descriptions apply to the next training round (not instantly)
Changing a topic label does not auto-update the description. Usually, this is not an issue if it is still referring to the same topic, but if you change it to something very different, you will have an outdated and irrelevant description.
To generate a new AI-written description, delete the current one (leave it blank). However, note that this will also only be re-generated on the next AI update.

Manual Reviewing Tips

It is still possible and useful to manually review rows! However, we expect that the new AI should have a solid enough baseline to reduce the amount of manual reviewing required.

Where to Focus Your Time

Update the descriptions and make topics clearer instead of reviewing many individual examples
If reviewing is necessary or you prefer it over editing descriptions:
- Skip easy/obvious rows, the AI got it right
- Focus on borderline or complex examples

Label fewer rows, but make them count.
(The AI now benefits more from quality over quantity.)

AI Updates

AI updates are no longer automatically triggered; use the button to trigger an AI update after meaningful changes. Use the AI runs sparingly as only a limited number is included for free for each analyzed column.

‼️ Note that this new behaviour is still highly work in progress, and we appreciate your feedback

As the new AI is much more performant, it is no longer necessary to run an AI update on every change of topic or category labels. Also, the new LLM is very expensive to run, so we have to limit the number of included runs. We have thus removed the automatic trigger to run the new AI.

Instead, we allow you as a user to choose when you want to run an AI update. Use the provided button to trigger an update if you made meaningful changes to the topics or have reviewed rows and want them to take effect.

Each open-ended column can be updated with AI a maximum of 3 times in total (including the first run). This limit applies for the lifetime of the project, not per time frame.
If you’d like to run additional updates beyond these 3, that’s possible — they would just consume extra credits.
The extra credit cost is 50% of the original import. For example, if the first import of a column used 1,000 credits, then each additional run would cost 500 credits.

This setup is part of the current beta phase, so things may evolve. The idea is that thanks to the high quality of the AI, multiple re-runs usually aren’t needed, but we’ll continue improving the flexibility going forward.

Troubleshooting topic assignment:

If you notice any of the following behaviors, please reach out to Support (support@caplena.com or in chat) for assistance:
- Too many topics assigned: This can happen when topics overlap or are very similar, which may confuse the AI.
- Batches of blank rows: In some cases, the AI may leave consecutive rows without any topics assigned.

Supported languages:

The new AI supports more languages natively

Afrikaans (af)
Albanian (sq)
Amharic (am)
Arabic (ar)
Armenian (hy)
Assamese (as)
Azerbaijani (az)
Basque (eu)
Belarusian (be)
Bengali (bn)
Bosnian (bs)
Bulgarian (bg)
Catalan (ca)
Cebuano (ceb)
Chinese (Simplified and Traditional) (zh)
Corsican (co)
Croatian (hr)
Czech (cs)
Danish (da)
Dhivehi (dv)
Dutch (nl)
English (en)
Esperanto (eo)
Estonian (et)
Filipino (Tagalog) (fil)
Finnish (fi)
French (fr)
Frisian (fy)
Galician (gl)
Georgian (ka)
German (de)
Greek (el)
Gujarati (gu)
Haitian Creole (ht)
Hausa (ha)
Hawaiian (haw)
Hebrew (iw)
Hindi (hi)
Hmong (hmn)
Hungarian (hu)
Icelandic (is)
Igbo (ig)
Indonesian (id)
Irish (ga)
Italian (it)
Japanese (ja)
Javanese (jv)
Kannada (kn)
Kazakh (kk)
Khmer (km)
Korean (ko)
Krio (kri)
Kurdish (ku)
Kyrgyz (ky)
Lao (lo)
Latin (la)
Latvian (lv)
Lithuanian (lt)
Luxembourgish (lb)
Macedonian (mk)
Malagasy (mg)
Malay (ms)
Malayalam (ml)
Maltese (mt)
Maori (mi)
Marathi (mr)
Meiteilon (Manipuri) (mni-Mtei)
Mongolian (mn)
Myanmar (Burmese) (my)
Nepali (ne)
Norwegian (no)
Nyanja (Chichewa) (ny)
Odia (Oriya) (or)
Pashto (ps)
Persian (fa)
Polish (pl)
Portuguese (pt)
Punjabi (pa)
Romanian (ro)
Russian (ru)
Samoan (sm)
Scots Gaelic (gd)
Serbian (sr)
Sesotho (st)
Shona (sn)
Sindhi (sd)
Sinhala (Sinhalese) (si)
Slovak (sk)
Slovenian (sl)
Somali (so)
Spanish (es)
Sundanese (su)
Swahili (sw)
Swedish (sv)
Tajik (tg)
Tamil (ta)
Telugu (te)
Thai (th)
Turkish (tr)
Ukrainian (uk)
Urdu (ur)
Uyghur (ug)
Uzbek (uz)
Vietnamese (vi)
Welsh (cy)
Xhosa (xh)
Yiddish (yi)
Yoruba (yo)
Zulu (zu)