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

3.0 KiB

Raw Permalink Blame History

last_updated, tags, source

last_updated

Voyage AI — voyage-4-large & voyage-4

Overview

Voyage AI's latest embedding models. Founded by OpenAI alumni. 200M free tokens — risk-free testing.

voyage-4-large (recommended)

Type: Cloud
Provider: Voyage AI
Price: $0.12 per 1M input tokens (200M free tokens included) (source)
Max input tokens: 32,000
Embedding dimensions: 1024
Language coverage: Multilingual
Persian benchmark: No public Persian-specific scores. MTEB multilingual aggregate available.
Notes: Slightly cheaper than OpenAI text-embedding-3-large ($0.12 vs $0.13). 32K context — 4× OpenAI's 8K. voyage-3-large and voyage-3 are now legacy.

voyage-4 (mid-tier)

Type: Cloud
Provider: Voyage AI
Price: $0.06 per 1M input tokens (200M free tokens included) (source)
Max input tokens: 32,000
Embedding dimensions: 1024
Language coverage: Multilingual
Persian benchmark: No public Persian-specific scores.

Pros

$0.12/1M tokens — slightly cheaper than OpenAI
200M tokens free — enough for testing + months of low-volume traffic
32K context — 4× OpenAI's context window
Strong MTEB multilingual scores
Simple API (OpenAI-compatible format)

Cons

No published Farsi-specific scores
Smaller company — less battle-tested at enterprise scale
Requires separate API key management

API Test — Curl

Endpoint: POST https://api.voyageai.com/v1/embeddings Auth: Authorization: Bearer $VOYAGE_API_KEY Get key: https://voyage.ai/api-keys Free tier: 200M tokens

voyage-4-large

curl -s https://api.voyageai.com/v1/embeddings   -H "Content-Type: application/json"   -H "Authorization: Bearer $VOYAGE_API_KEY"   -d '{
    "model": "voyage-4-large",
    "input": ["بیمه نامه شخص ثالث چیست؟"]
  }' | python3 -m json.tool

voyage-4

curl -s https://api.voyageai.com/v1/embeddings   -H "Content-Type: application/json"   -H "Authorization: Bearer $VOYAGE_API_KEY"   -d '{
    "model": "voyage-4",
    "input": ["بیمه نامه شخص ثالث چیست؟"]
  }' | python3 -m json.tool

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

API Test — Python

import voyageai

vo = voyageai.Client(api_key="YOUR_VOYAGE_API_KEY")  # or uses VOYAGE_API_KEY env var

result = vo.embed(
    ["بیمه نامه شخص ثالث چیست؟"],
    model="voyage-4-large",
    input_type="document"
)

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

TODO

Run 100-question eval set against voyage-4-large (200M free tokens)
Benchmark on MIRACL-Farsi subset
Compare with OpenAI text-embedding-3-large on Farsi queries
Calculate cost at production volume

3.0 KiB Raw Permalink Blame History Unescape Escape