Diba-Embed

English

Diba-Embed is the first text-embedding model from Dibachain and the first embedding model in the Diba family — a Persian-first model that turns Persian and English text into dense vectors, placing similar meanings close together. Built by Dibachain (dibachain.ir) and tuned and packaged for Iranian, Persian-language use cases where most open models fall short.

Use it to build semantic search, retrieval‑augmented generation (RAG), FAQ matching, clustering, deduplication, and reranking — in Persian, in English, and across the two.

What it can do

• Persian semantic search — find the most relevant document for a Persian question, even when the wording differs.
• Retrieval for RAG — retrieve the right passages to ground a chat model such as Diba-Base on your own documents.
• Cross‑lingual matching — a Persian query can find an English passage and vice‑versa.
• Clustering & deduplication — group similar tickets, comments, or products; detect near‑duplicates.
• Reranking — order candidate passages by relevance to a query.

Specifications

Persian retrieval benchmark

Measured on 400 Persian question/answer pairs drawn from Diba's own grounded data (Iran history, contemporary topics, and Dibachain company Q/A). Each question must retrieve its correct passage out of the full pool. Greedy, on CPU, no task‑specific training:

In 95% of cases the correct passage is among the top three results — enough to drive reliable Persian RAG and search.

Quick start

Diba-Embed ships with the Diba model definition, so pass trust_remote_code=True when loading.

sentence-transformers (recommended):

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Dibachain/Diba-Embed", trust_remote_code=True)

# queries use the built-in "query" prompt; documents are embedded as-is
queries = ["پایتخت ایران کجاست؟"]
documents = [
    "تهران پایتخت و بزرگ‌ترین شهر ایران است.",
    "اصفهان یکی از شهرهای تاریخی ایران است.",
]

q = model.encode(queries, prompt_name="query", normalize_embeddings=True)
d = model.encode(documents, normalize_embeddings=True)
scores = q @ d.T
print(scores)   # highest score points to the matching document

transformers (last-token pooling, for full control):

import torch, torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Dibachain/Diba-Embed", padding_side="left")
model = AutoModel.from_pretrained("Dibachain/Diba-Embed", trust_remote_code=True).eval()

def embed(texts, is_query=False):
    if is_query:
        texts = [f"Instruct: Given a query, retrieve passages that answer it\nQuery: {t}" for t in texts]
    enc = tok(texts, padding=True, truncation=True, max_length=512, return_tensors="pt")
    with torch.no_grad():
        h = model(**enc).last_hidden_state
    lengths = enc["attention_mask"].sum(1) - 1
    v = h[torch.arange(h.size(0)), lengths]         # last non-pad token
    return F.normalize(v, dim=1)

sim = embed(["پایتخت ایران؟"], is_query=True) @ embed(["تهران پایتخت ایران است."]).T
print(sim)

Tip: always add the query instruction to search queries; embed documents without it. Normalize vectors and rank by cosine (dot product of normalized vectors).

Run with llama.cpp (GGUF)

Diba-Embed ships ready-to-run llama.cpp files (GGUF format; the .bin files load directly with -m) for CPU inference with llama.cpp — no Python required.

# start an OpenAI-compatible embeddings server on CPU
llama-server -m diba-embed-q8_0.bin --embedding --pooling last -c 2048 --port 8080

# then request embeddings
curl http://localhost:8080/v1/embeddings   -H "Content-Type: application/json"   -d '{"input": ["تهران پایتخت ایران است", "capital of Iran"]}'

Direct download: https://huggingface.co/Dibachain/Diba-Embed/resolve/main/diba-embed-q8_0.bin

Intended use and limitations

Diba-Embed is built for search and retrieval, not for generation — it produces vectors, not text. Quality is strongest on general and formal Persian and on the domains in the benchmark above; very specialized or noisy text may need domain adaptation. It reflects biases present in its training data. For generation and chat, use Diba-Base.

فارسی

دیبا-امبد نخستین مدل بردارسازی متن (Text Embedding) شرکت دیباچین و اولین مدل امبدینگ خانواده‌ی دیبا است — مدلی فارسی‌محور که متن فارسی و انگلیسی را به بردارهای عددی تبدیل می‌کند تا متن‌های هم‌معنا در فضای برداری به هم نزدیک شوند. ساخته‌ی شرکت دیباچین (dibachain.ir) و ویژه‌ی کاربردهای فارسی‌زبان و ایرانی، جایی که