ajan-embed-q

A Turkish-optimized sentence embedding model for retrieval / RAG, distilled from BAAI/bge-m3 into a multilingual-e5-base student over 500k Turkish web sentences. 768→1024-dim projected to match the teacher's space.

🔗 Code + recipe: github.com/AJANLAR-AI/ajanlar

Part of Ajanlar — open Turkish AI agent infra. The retrieval engine under the agents, where multilingual models underperform on agglutinative Turkish.

Benchmark (MTEB, Turkish) — with ablation

Main score per task (NDCG@10 retrieval; Spearman STS). multilingual-e5-base is the undistilled ablation (same base as this model).

Model	Params	TurHistQuad (retrieval)	STS22.v2	STS17	avg
ajan-embed-q (this)	278M	0.465	0.651	0.724	0.613
multilingual-e5-small	118M	0.433	0.643	0.767	0.614
multilingual-e5-base (ablation)	278M	0.444	0.651	0.777	0.624
multilingual-e5-large	560M	0.469	0.675	0.810	0.652
BAAI/bge-m3 (teacher)	568M	0.478	0.680	0.814	0.657

Honest reading: this is a retrieval-specialised model. On Turkish retrieval (TurHistQuad) it beats e5-small/e5-base and matches e5-large at half the size — the RAG/agent use case it's built for. On general STS it trails the e5 family, so on the 4-task average it lands ~on par with (slightly below) undistilled e5-base. Use it for retrieval/RAG in Turkish, not as a general-purpose STS model.

Usage

from sentence_transformers import SentenceTransformer
m = SentenceTransformer("fredoline005/ajan-embed-q")
emb = m.encode(["query: kargom ne zaman gelir?",
                "passage: Siparişler 1–3 iş günü içinde kargoya verilir."],
               normalize_embeddings=True)

Use query: / passage: prefixes for retrieval (inherited from the e5 family).

Training

• Method: offline distillation — teacher embeddings cached over the corpus, the student trained (MSE) to reproduce them; a Dense layer projects 768→1024.
• Teacher: BAAI/bge-m3 (MIT). Student: intfloat/multilingual-e5-base (MIT).
• Data: 500k Turkish sentences (allenai/c4, tr).
• Hardware: 1× RTX 4090, ~1 hour.

Limitations (honest)

• Benchmarked on 2 Turkish tasks, no confidence intervals — margins are modest.
• The win over e5-small is partly the larger base. The undistilled-e5-base ablation HAS now been run and is in the table above: it beats this model on the 4-task average (0.624 vs 0.613) and loses to it on retrieval (0.444 vs 0.465). Distillation bought retrieval and cost general STS.
• Distillation is bounded by the teacher; the native-Turkish-tokenizer edge is v1.
• Pin a corpus snapshot + add PII/dedup filtering for a production retrain.

License

Apache-2.0 (weights/recipe). Base + teacher are MIT.

Ajanlar is available for acquisition

The Ajanlar project is for sale as a whole: the ajanlar.ai domain, this model and its training recipe, and the brand.

It was built because multilingual embedding models underperform on agglutinative Turkish. A 278M-parameter distilled student closes most of that gap on retrieval and matches multilingual-e5-large at half the size on TurHistQuad. The model is installed roughly 240 times a month and that rate has about doubled since June with no promotion of any kind.

This changes nothing for anyone using the model. It stays Apache-2.0, it stays up, and it stays free, whoever ends up owning it. If you are building on it, keep building.

Enquiries: support@ranked.ng

Who made this

Built by Ranked Technologies Limited (RC 9522220) as the retrieval engine under Ajanlar, Turkish-language AI support agents that answer from a business's own documents.

If you are using this model, we would genuinely like to know what for. If it is falling short on your data, tell us and we will look at it.

Contact: support@ranked.ng

We also build Nigerian-language AI at 9jatesters and publish speech datasets and an open benchmark at 9jaBench.