Qaf · Research status labels · Oct 22, 2026Research labels are back to the GPT-5.6 style
The “at most 5 words” rule made Gemini stack keywords (“Dua for physical pain Uthman”). The new wording asks for a short phrase with normal grammar, and its labels look like the ones GPT-5.6 sol wrote.
Shape of the labels (English UI)
| Median words | Starts with “Search” | Starts with a verb | Example | |
|---|---|---|---|---|
| GPT-5.6 sol (prod, Jul 25–Aug 14) | 5 | 1% | 46% | Verify Hanbali endowment evidence |
| Gemini, original wording (prod) | 9 | 30% | 31% | Search ruling on pledging allegiance to rulers in democratic or non-Sharia systems |
| Gemini, “at most 5 words” (prod now) | 5 | 24% | 24% | search istef'al transitive intransitive |
| Gemini, new wording (eval) | 5–6 | 0–7% | 30–48% | Ruling on dragging pants below ankles |
Production rows: about 4,000 labels sampled from Langfuse. Eval rows: 80 real questions replayed through Gemini 3.7 Flash with the production prompt, with only the label instructions changed (ranges cover two runs and both thinking levels).
Blind comparison, new vs current
88% / 84%Better status line (English / Arabic UI), share of decided comparisons
87% / 76%Closer to GPT-5.6 style (English / Arabic UI); about half the Arabic pairs were ties
16 AI judges, two per batch with different focus. Each saw the question and both label sets in random A/B order, with real GPT-5.6 labels as the style reference. These are AI judgements, not human ratings.
What GPT-5.6 sol wrote
The prompt change
BEFORE · platform/main-v2 v4
**Tool labels**
For every `search` and `lookup_catalog` call, set `label` to a concise description of the intent in at most 5 words, written in {{language}}. Shorter labels are welcome; there is no minimum word count.AFTER
**Tool labels**
Every `search` and `lookup_catalog` call needs a `label`: a short status line the user sees while you research. Write it in {{language}}, even when the query is in Arabic.
- Say what this call is looking for in a natural phrase of 3–5 words (6 at most), the way a researcher notes a step: "Find juristic views on timing", "Verify Hanbali endowment evidence", "Ruling on income from music", "Four schools on eyebrow plucking", "Identify Ibn Abd al-Barr", "تخريج حديث أعلنوا النكاح", "بيان الفرق بين النبي والرسول".
- Keep normal grammar with small words like "on", "of", "for"; never stack keywords (not "Dua physical pain Uthman" but "Find Uthman's dua for pain").
- Never start with "Search", "Look up" or "البحث عن". Do not paste a verse, hadith, dua or the query into the label; name it instead ("Find the hadith on three qirats").
- Sentence case, no quotes, no ending punctuation. Give each call in a turn its own specific label.The tool schema's label description is updated to match in the PR.