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

3.5 KiB

Raw Permalink Blame History

last_updated, tags, source

last_updated

Jina AI — jina-embeddings-v3

Overview

Available as both cloud API and local model. Cheapest cloud option with explicit Farsi support. ⚠️ Local model has non-commercial license (CC-BY-NC-4.0).

Cloud API

Type: Cloud
Provider: Jina AI
Price: ~$0.02 per 1M tokens (source)
Max input tokens: 8,192
Embedding dimensions: 1024 (Matryoshka — resizable)
Language coverage: 89 languages including Farsi
Persian benchmark: No public Persian-specific scores. Strong MTEB multilingual.

Local Model

Parameters: ~568M
VRAM (fp16): ~1.1 GB
License: CC-BY-NC-4.0 ⚠️ (non-commercial!) — verify for enterprise use
Model card: https://huggingface.co/jinaai/jina-embeddings-v3

Pros

Cheapest cloud option: ~$0.02/1M tokens
89 languages including Farsi
Task-specific adapters (retrieval, clustering, classification)
Matryoshka support (resizable dimensions)
8K context
ColBERT/late interaction support

Cons

Local model is CC-BY-NC (non-commercial) — verify for enterprise
Smaller company — less enterprise track record
Pricing page not transparent (SPA, not standard display)

API Test — Curl

Endpoint: POST https://api.jina.ai/v1/embeddings Auth: Authorization: Bearer $JINA_API_KEY Get key: https://jina.ai/api-keys

For queries

curl -s https://api.jina.ai/v1/embeddings   -H "Content-Type: application/json"   -H "Authorization: Bearer $JINA_API_KEY"   -d '{
    "model": "jina-embeddings-v3",
    "input": ["بیمه نامه شخص ثالث چیست؟"],
    "task": "retrieval.query"
  }' | python3 -m json.tool

For document chunks (indexing)

curl -s https://api.jina.ai/v1/embeddings   -H "Content-Type: application/json"   -H "Authorization: Bearer $JINA_API_KEY"   -d '{
    "model": "jina-embeddings-v3",
    "input": ["بیمه نامه شخص ثالث شامل پوشش خسارات مالی و جانی است."],
    "task": "retrieval.passage"
  }' | python3 -m json.tool

With Matryoshka dimension reduction

curl -s https://api.jina.ai/v1/embeddings   -H "Content-Type: application/json"   -H "Authorization: Bearer $JINA_API_KEY"   -d '{
    "model": "jina-embeddings-v3",
    "input": ["بیمه نامه شخص ثالث چیست؟"],
    "task": "retrieval.query",
    "dimensions": 512
  }' | python3 -c "import sys,json; d=json.load(sys.stdin); print(f'dims={len(d["data"][0]["embedding"])}')"

Key differences from OpenAI:

Use task: "retrieval.query" for queries, task: "retrieval.passage" for chunks
Optional dimensions param for Matryoshka truncation
Optional late_chunking: true for late-interaction-style results

Response shape: data[0].embedding = float array, usage.total_tokens = token count

API Test — Python

from jina import Client

client = Client(api_key="YOUR_JINA_API_KEY")  # or uses JINA_API_KEY env var

result = client.embed(
    inputs=["بیمه نامه شخص ثالث چیست?"],
    model="jina-embeddings-v3",
    task="retrieval.query"
)

embedding = result.embeddings[0].embedding
print(f"Dimensions: {len(embedding)}")

TODO

Verify enterprise licensing terms for local model (CC-BY-NC)
Test cloud API with 100-question eval set
Benchmark on MIRACL-Farsi subset
Compare with OpenAI and Cohere on Farsi queries

3.5 KiB Raw Permalink Blame History