Mahdi Bazrafshan
6c9cebaaa0
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)
3.5 KiB
3.5 KiB
last_updated, tags, source
| last_updated | tags | source | |||||||
|---|---|---|---|---|---|---|---|---|---|
| 2026-07-20 |
|
import-knowledge |
Google Vertex AI — Embedding Models
Models
Gemini Embedding (newest)
- Type: Cloud
- Provider: Google Vertex AI
- Price: $0.15 per 1M input tokens (online) / $0.12 per 1M (batch) (source)
- Max input tokens: 8,192
- Embedding dimensions: 768
- Language coverage: Multilingual (Gemini family)
text-embedding-004
- Price: ~$0.10 per 1M tokens equivalent ($0.025 per 1,000 characters) (source)
- Language coverage: English-primary with some multilingual
text-multilingual-embedding-002 (Farsi-relevant)
- Price: ~$0.10 per 1M tokens equivalent (same pricing as above)
- Language coverage: Multilingual — this is the relevant variant for Farsi
Note: Legacy Google embedding models priced per character, not per token (~4 chars/token).
Pros
- Google ecosystem integration
- Batch pricing available (20% discount)
- Gemini Embedding is Google's latest
Cons
- Most complex setup (requires GCP project, billing, Vertex AI API)
- No Farsi-specific scores published
- Per-character pricing for legacy models (confusing)
- $0.15/1M tokens for Gemini Embedding — most expensive option
API Test — Curl
Auth: gcloud auth print-access-token or service account JSON
Setup: Requires GCP project with Vertex AI API enabled
Gemini Embedding
export GCP_TOKEN=$(gcloud auth print-access-token)
curl -s "https://us-central1-aiplatform.googleapis.com/v1/projects/YOUR_PROJECT_ID/locations/us-central1/publishers/google/models/gemini-embedding-exp-03-07:predict" -H "Content-Type: application/json" -H "Authorization: Bearer $GCP_TOKEN" -d '{
"instances": [
{"content": "بیمه نامه شخص ثالث چیست؟"}
]
}' | python3 -m json.tool
text-multilingual-embedding-002
curl -s "https://us-central1-aiplatform.googleapis.com/v1/projects/YOUR_PROJECT_ID/locations/us-central1/publishers/google/models/text-multilingual-embedding-002:predict" -H "Content-Type: application/json" -H "Authorization: Bearer $GCP_TOKEN" -d '{
"instances": [
{"content": "بیمه نامه شخص ثالث چیست؟"}
]
}' | python3 -m json.tool
text-embedding-004
curl -s "https://us-central1-aiplatform.googleapis.com/v1/projects/YOUR_PROJECT_ID/locations/us-central1/publishers/google/models/text-embedding-004:predict" -H "Content-Type: application/json" -H "Authorization: Bearer $GCP_TOKEN" -d '{
"instances": [
{"content": "بیمه نامه شخص ثالث چیست؟"}
]
}' | python3 -m json.tool
Response shape: predictions[0].embeddings.values = float array (different from OpenAI!)
API Test — Python
from google.cloud import aiplatform
aiplatform.init(project="YOUR_PROJECT_ID", location="us-central1")
model = aiplatform.Endpoint("projects/YOUR_PROJECT_ID/locations/us-central1/endpoints/ENDPOINT_ID")
response = model.predict(instances=[{"content": "بیمه نامه شخص ثالث چیست؟"}])
embedding = response.predictions[0]["embeddings"]["values"]
print(f"Dimensions: {len(embedding)}")
TODO
- Set up GCP project if not already available
- Benchmark text-multilingual-embedding-002 on Farsi queries
- Compare with OpenAI on Farsi retrieval quality