Files

Mahdi Bazrafshan 6c9cebaaa0 docs(research): add farsi embedding model research knowledge base

Body:
        Why:
        - Insurance RAG pipeline needs Farsi-optimized embeddings
        - Current local baseline (Nomic Embed) is English-only

        Changes:
        - 8 cloud model profiles with curl + python test commands
        - 5 local model profiles with VRAM, serving platform notes
        - Persian benchmark landscape (MIRACL-Fa, mMARCO, FaMTE, MTEB)
        - Decision report with ranked options and recommendations
        - Full comparison table (15 models × 10 columns)

        Impact:
        - Knowledge base for embedding model selection
        - Top picks: BGE-M3 (local), Cohere embed-v4.0 (cloud)
        - Critical: Nomic Embed must be replaced (English-only)

2026-07-20 15:00:11 +03:30

2.2 KiB

Raw Blame History

last_updated, tags, source

last_updated

Alibaba-NLP/gte-multilingual-base

Overview

Type: Local
Provider: Alibaba NLP
Parameters: ~305M
VRAM (fp16): ~610 MB
Embedding dimensions: 768
Max input tokens: 8,192
Language coverage: 70+ languages explicitly including Farsi (fa)
Serving platform: sentence-transformers, TEI
License: Apache-2.0
Model card: https://huggingface.co/Alibaba-NLP/gte-multilingual-base

Why Consider It

Lightest model with confirmed Farsi support. Only 610 MB VRAM. 8K context. Apache-2.0 license. Benchmarked on MIRACL and PL-MTEB. Simpler than BGE-M3 (dense-only, no ColBERT/sparse overhead).

Pros

Explicitly lists Farsi in supported languages
Only 610 MB VRAM — lightest meaningful option with Farsi
8,192 token context
Benchmarked on MIRACL and PL-MTEB
Apache-2.0 license
Simpler architecture (dense-only) — easier to deploy and debug

Cons

No ColBERT/sparse support — dense-only retrieval
No published isolated Farsi retrieval scores
Alibaba NLP ecosystem less mature than HuggingFace ecosystem
Not available on Ollama (may need manual GGUF conversion)

API Test — Python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Alibaba-NLP/gte-multilingual-base")

# Farsi insurance query
query = "شرایط لغو بیمه نامه چیست？"
embedding = model.encode(query)
print(f"Dimensions: {len(embedding)}")
# Expected: Dimensions: 768

API Test — TEI (Docker)

# Start TEI server
docker run --gpus all -p 8080:80   -v $PWD/data:/data   ghcr.io/huggingface/text-embeddings-inference:latest   --model-id Alibaba-NLP/gte-multilingual-base

# Test
curl -s http://localhost:8080/embed   -H "Content-Type: application/json"   -d '{"inputs": "بیمه نامه شخص ثالث چیست؟"}'   | python3 -c "import sys,json; d=json.load(sys.stdin); print(f'dims={len(d[0])}')"

TODO

Benchmark on MIRACL-Farsi subset
Compare with BGE-M3 on Farsi retrieval quality
Test GGUF conversion for Ollama compatibility
Measure inference latency on RTX 3090

2.2 KiB Raw Blame History