Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Mir Kvestov Search Parses Plain-Language Queries: 37% of Searches Reach a Quest Page

A Mir Kvestov visitor types "something creepy for two," but the catalog has no such category, and search could not understand the query. Now six LLM classifiers turn the phrase into filters: fear level 2–3 and two players. As a result, 37% of searches end with the visitor opening a quest page, another 52% end with filter adjustments, and only 11% end with no action. If you have no user history, a buyer survey and an LLM are enough.

AI-processed from Habr AI; edited by Hamidun News
Mir Kvestov Search Parses Plain-Language Queries: 37% of Searches Reach a Quest Page
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Post-booking questionnaire10–12 questions, 20–30 thousand responses a year
Key metric #2A depth-6 tree compressed 120 checkboxes into 3–5 clicks, with 76% precision
Key metric #365% of visitors are not logged in, so there is no browsing history for them

Imagine a visitor to a site with three thousand quests typing "something creepy for two" into the search bar. The catalog has no such category. It has fear levels from 0 to 3, player counts, formats and themes.

A search that matches exact words will return nothing or a random list, and the person will leave without booking. The quest aggregator "Mir Kvestov" (7.5 million visitors a year) solved exactly this problem: six language models running in classification mode (LLM classifiers) translate a free-form phrase into catalog filters.

As a result, 37% of searches end with the person opening a quest page, 52% end with them changing filters and looking at the results again, and only 11% lead to neither. The story is useful to developers of niche catalogs because it shows how to get by without large user logs.

The problem: nobody to recommend to, and nothing to recommend from

The quest market is mature and demand, according to Wordstat, is flat, so the main growth reserve is helping people who already arrived find their quest. But 65% of visitors are not logged in. There is no browsing history for them, and a classic recommender system works precisely from history. The default results showed everyone the popular quests: scary ones and performance-style ones. That did not suit people who wanted an adventure or a detective story; by the author's estimate, they make up 11–17% of visitors.

Instead of history, the team asked the buyers themselves. After a booking, the site offers a questionnaire of 10–12 questions: genres, settings, fear and difficulty levels, who the person plays with and on what occasion. Over a year it collects 20–30 thousand questionnaires, each tagged with the quest that was booked. The result is a labeled dataset: preferences on the left, the choice on the right.

Version 1: 120 checkboxes become three clicks

The questionnaire has about 120 answer options, and nobody will tick them all on a website. So the team compressed the questions into three steps. The answers were encoded as 0/1 features. Quests were grouped with k-means into classes based on the average profile of their buyers. Then a decision tree of depth 6 was trained to guess a quest's class from a minimum of answers.

On the site, this appears as suggestion buttons under the search bar. The visitor taps "Not scary," "For kids," "magical creatures," and the phrase assembles itself in the search bar: "A not-scary quest for kids about magical creatures in the action genre for 4 players." The tree's precision is 76% (the share of correct predictions), recall is 65% (how many of the correct classes the model finds), and the F1 score is 67% (the harmonic mean of the two). Three clicks are enough for a recommendation, and people usually click 4–5 times.

The downside showed up quickly: the categories were invented by professionals, and a visitor is under no obligation to tell suspense from thriller and horror. People who clicked the suggestions got a recommendation; people who typed did not.

Version 2: six classifiers parse the text on the fly

The team then added six LLM classifiers, one per parameter: fear level (0–3), difficulty, number of players, minimum age, format and theme. The query goes to the models while the person is still typing. The phrase "scary for two" becomes fear 2–3 and two players, and the other fields stay empty.

Two things happen next. The suggestion tree switches by itself to the "Scary or horror" branch, so the visitor sees that the system understood them. The recognized words become exact filters, and only the remainder, such as a quest title, goes to the text search. Input is hybrid: you can click the "Mystery" chip and add "for two," or start with text and refine with clicks. A chip is removed with one tap; text is deleted character by character.

The metrics and how to read them

A success was opening a quest page after a search; a reformulation was a filter change without opening a page. A pause of more than 30 minutes starts a new session. A reformulation is not necessarily a failure: the person looked at the results and is fine-tuning the query. These are product metrics, not booking conversion.

For reference, the author cites public data: in Yandex studies from 2009–2014, 42–49% of sessions contained a repeated query, and at eBay more than half of queries are part of reformulation chains (Hirsch et al., SIGIR 2020). The sources count different things (searches, sessions, queries), so only orders of magnitude can be compared, not exact percentages. By order of magnitude, Mir Kvestov is in the same range as large search engines.

The giants use similar techniques. In 2014, Pinterest added clickable refinements and got 25% more searches per person. In 2025, Instacart serves frequent queries from a cache and rare ones with a fine-tuned Llama 3 8B in about 300 ms; complaints about poor results fell by half. In 2026, Airbnb solved its cold start with about 500 user surveys and synthetic queries generated by an LLM. The difference is in the data: tech giants have millions of products and logs, while a niche catalog has few, so here the team went with a survey and a tree.

What to take away

If you have no user data, ask the people who have already bought: a survey gives you a labeled dataset. A long questionnaire can be compressed with a tree: 120 checkboxes became 3–5 clicks. And do not make customers think in the language of your categories: let them write what they want, and the classifiers will translate the text into filters. The hybrid search is live on mir-kvestov.ru.

Story

Key Questions Answered

What are the key details on The problem: nobody to recommend to, and nothing to recommend from?

The quest market is mature and demand, according to Wordstat, is flat, so the main growth reserve is helping people who already arrived find their quest. But 65% of visitors are not logged in. There is no browsing history for them, and a classic recommender system works precisely from history. The default results showed everyone the popular quests: scary ones and performance-style ones. That did not suit people who wanted an adventure or a detective story; by the author's estimate, they make up 11–17% of visitors.

What are the key details on Version 1: 120 checkboxes become three clicks?

The questionnaire has about 120 answer options, and nobody will tick them all on a website. So the team compressed the questions into three steps. The answers were encoded as 0/1 features. Quests were grouped with k-means into classes based on the average profile of their buyers. Then a decision tree of depth 6 was trained to guess a quest's class from a minimum of answers. On the site, this appears as suggestion buttons under the search bar.

What are the key details on Version 2: six classifiers parse the text on the fly?

The team then added six LLM classifiers, one per parameter: fear level (0–3), difficulty, number of players, minimum age, format and theme. The query goes to the models while the person is still typing. The phrase "scary for two" becomes fear 2–3 and two players, and the other fields stay empty. Two things happen next. The suggestion tree switches by itself to the "Scary or horror" branch, so the visitor sees that the system understood them.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…