This is a synthetic benchmark for searching team-chat messages in English, Brazilian Portuguese, Swedish and Japanese. All people, companies and events in it are fictional. Please tell us a little about yourself and how you plan to use it.

Kestrel Cloud: Multilingual Chat Message Search

A benchmark for finding messages in a team chat when people write in several languages, often within the same message. It contains a synthetic chat workspace (6,014 messages), 360 search queries written the way people type into a chat search box, and graded relevance labels for 22,480 query–message pairs.

It was built to choose a search configuration for Rocket.Chat message search: keyword search (PostgreSQL full-text search and BM25), semantic (vector) search, and hybrids of the two.

What's inside


Workspace	"Kestrel Cloud", a fictional B2B data company with offices in São Paulo, Stockholm and Tokyo plus a remote English-speaking team
Messages	6,014 across 27 channels and 33 users, January–October 2026
Languages	English 2,136; Brazilian Portuguese 1,365; Swedish 1,284; Japanese 1,213; three-language 16. 2,082 messages mix languages inside the message.
Queries	360 in 17 categories (below), split into 108 dev and 252 test queries, stratified by category
Relevance labels	22,480 graded pairs (0–3): 5,210 planted by the dataset generator, 17,270 added by blind pooled judging

Messages look like real chat: short and informal, with typos, missing accents, emoji, @mentions, links, code in backticks, bot alerts, threads, and about 10% long messages such as logs and announcements. Lookalike messages ("hard negatives") that share words with a query but are about something else are planted on purpose.

Query categories

Category	Example query → message it should find
`typo`	`kuberentes 1.32 upgrde` → "Kubernetes 1.32 upgrade plan for the EKS clusters: …"
`code_mixed`	`subi hotfix billing-svc` → "subi hotfix billing-svc v4.14.3 com checagem da idempotency key …"
`script_variant`	`ＫＩＴＥ－８４２` (full-width) → "KITE-842 created: iPad の設定画面でレイアウトが崩れる …"
`cross_lingual`	`tax invoice issued twice for the same order at the online store` → a Portuguese message about a duplicated NF-e
`paraphrase_nl`	`why did clients receive the same report two times` → "export-scheduler cron fired 2x after the leader election glitch …"
`identifier`	`SUP-1342` → "Banco Horizonte opened SUP-1342: the reports in Kestrel Lens are very slow …"

The other categories are proper_noun, keyword_known_item, exact_phrase, morphology, short_broad, native_monolingual, acronym_abbrev, numeric_temporal, negation_constraint, mixed_message (a mixed-language message searched in one language) and long_message (effects of message length).

Files and fields

corpus (data/corpus.jsonl): one message per line.

Field	Description
`msg_id`	Message id, e.g. `P06-0001`
`room_name`, `room_id`	Channel name and a random id
`username`, `user_id`	Fictional user and a random id
`ts`	ISO 8601 timestamp (UTC)
`thread_id`	`msg_id` of the thread parent, or `null`
`lang`	`en`, `pt`, `sv`, `ja`, `pt+en`, `sv+en`, `ja+en` or `multi`
`text`	Message text

queries (data/queries-{dev,test}.jsonl): query_id, text, category, query_lang, intent (a precise description of the information need) and hard_negatives (lookalike msg_ids planted for this query).

qrels (data/qrels-{dev,test}.jsonl): query_id, msg_id, grade and label_source.

Grade	Meaning
3	The message the user is looking for, or one that directly answers the need
2	About the same specific thing (incident, ticket, decision) and useful
1	Same general topic, would not really help
0	Not relevant, including deliberate lookalikes

label_source is generator for labels written together with the query, or pooled_judge for labels added later by blind judging. Messages without a label are unjudged and are usually treated as not relevant. Pooling covered the full top 10 of every system in the baseline table below; a new system may surface unjudged messages, so report how many of its top 10 are judged.

from datasets import load_dataset

repo = "dnouv/kestrel-multilingual-chat-search"
corpus = load_dataset(repo, "corpus", split="corpus")
queries = load_dataset(repo, "queries", split="test")
qrels = load_dataset(repo, "qrels", split="test")

How it was built

Specification. A shared brief fixed the company, people, products, identifier formats, languages and realism rules, and the 17 query categories.
Generation, query first. Language-model agents (Anthropic Claude) each wrote one "pack" of the workspace. For every query they first decided the information need, then wrote the target messages, related messages and lookalikes, and labelled every message in their pack.
Pooling and blind judging. Every search system under evaluation contributed its top 10 results (the main systems their top 20). Unlabelled results were judged blind by language-model assessors against fixed guidelines, with the query intent and thread context shown and the system hidden.
Quality check. Judges re-labelled 815 hidden generator-labelled pairs. Agreement: Cohen's κ = 0.75 for relevant vs not relevant (grade ≥2), 69% exact grade match, 99.3% within one grade.

Baseline results

Test split (252 queries); relevant means grade ≥2 for MRR and recall. Systems searched identical data on PostgreSQL 18 (vchord_bm25 0.3.0, pg_tokenizer 0.1.1, VectorChord 1.1.1); dense embeddings came from ONNX MiniLM models. Differences of about 0.03 nDCG@10 or less are within noise.

System	nDCG@10	MRR@10	Recall@50
MongoDB `$text`, sorted newest first	0.131	0.167	0.228
MongoDB `$text`, sorted by text score	0.345	0.523	0.390
PostgreSQL full-text search (`ts_rank`)	0.214	0.388	0.166
BM25, English BERT tokenizer	0.406	0.590	0.486
BM25, XLM-R tokenizer, lowercased + NFKC	0.437	0.620	0.488
BM25, multilingual BERT (uncased) tokenizer	0.442	0.628	0.511
Dense, `all-MiniLM-L6-v2`	0.301	0.468	0.397
Dense, `paraphrase-multilingual-MiniLM-L12-v2`	0.251	0.385	0.345
Weighted RRF, dense + full-text (weight 0.5)	0.339	0.514	0.409
RRF, dense + full-text + BM25 (multilingual BERT)	0.442	0.640	0.545

Every system is weak on paraphrase_nl and cross_lingual queries (nDCG@10 ≤ 0.21), which leaves room for stronger multilingual embedding models.

Limitations

The messages are synthetic. They follow realistic chat patterns but are not real conversations.
The labels were made by language models. They were checked against hidden re-judgments, not by human assessors.
The corpus is small (6,014 messages), so it measures ranking quality, not behaviour at scale.
There is no recency or personalisation signal. Each query is judged purely on content.

License

CC BY 4.0. All names, companies, products, identifiers and events are fictional; any resemblance to real entities is coincidental.

Downloads last month: 10

Total file size:

4.26 MB