openova/platform/bge

BGE

BAAI General Embedding models for text embeddings and reranking. Application Blueprint (see docs/PLATFORM-TECH-STACK.md §4.6). Used by bp-cortex for embedding generation (Milvus vector search) and reranking (improving retrieval precision).

Status: Accepted | Updated: 2026-04-27

---

Overview

BGE provides state-of-the-art text embeddings and reranking for RAG systems, supporting multilingual (including Arabic) and hybrid sparse+dense retrieval.

flowchart LR
    subgraph BGE["BGE Services"]
        M3[BGE-M3<br/>Embeddings]
        Reranker[BGE-Reranker<br/>Cross-Encoder]
    end

    Text[Text] --> M3
    M3 --> Dense[Dense Vector<br/>1024-dim]
    M3 --> Sparse[Sparse Vector]

    Candidates[Candidates] --> Reranker
    Query[Query] --> Reranker
    Reranker --> Ranked[Ranked Results]

---

Models

Model	Purpose	Dimensions
BGE-M3	Multilingual embeddings	1024 (dense) + sparse
BGE-Reranker-v2-M3	Cross-encoder reranking	-

---

Why BGE?

Feature	Benefit
Multilingual	Arabic + English support
Hybrid retrieval	Dense + sparse vectors
High accuracy	MTEB benchmark leader
Reranking	Precision improvement
Self-hosted	Data sovereignty

---

Deployment

BGE-M3 Embeddings

apiVersion: apps/v1
kind: Deployment
metadata:
  name: bge-m3
  namespace: ai-hub
spec:
  replicas: 1
  template:
    spec:
      containers:
        - name: bge-m3
          image: harbor.<location-code>.<sovereign-domain>/ai-hub/bge-m3:latest
          ports:
            - containerPort: 8080
          env:
            - name: MODEL_ID
              value: "BAAI/bge-m3"
          resources:
            requests:
              cpu: "2"
              memory: 4Gi
              nvidia.com/gpu: "1"

BGE Reranker

apiVersion: apps/v1
kind: Deployment
metadata:
  name: bge-reranker
  namespace: ai-hub
spec:
  replicas: 1
  template:
    spec:
      containers:
        - name: bge-reranker
          image: harbor.<location-code>.<sovereign-domain>/ai-hub/bge-reranker:latest
          ports:
            - containerPort: 8080
          env:
            - name: MODEL_ID
              value: "BAAI/bge-reranker-v2-m3"
          resources:
            requests:
              cpu: "1"
              memory: 2Gi
              nvidia.com/gpu: "1"

---

API Endpoints

Embeddings

curl -X POST http://bge-m3.ai-hub.svc:8080/embed \
  -H "Content-Type: application/json" \
  -d '{
    "texts": ["Hello world", "مرحبا بالعالم"],
    "return_sparse": true
  }'

Response:

{
  "dense": [[0.123, 0.456, ...], [0.789, 0.012, ...]],
  "sparse": [{"token_id": 1234, "weight": 0.5}, ...]
}

Reranking

curl -X POST http://bge-reranker.ai-hub.svc:8080/rerank \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What is AML?",
    "documents": [
      "Anti-Money Laundering guidelines...",
      "Machine Learning tutorial...",
      "AML compliance requirements..."
    ]
  }'

Response:

{
  "scores": [0.95, 0.12, 0.89],
  "ranked_indices": [0, 2, 1]
}

---

LangChain Integration

from langchain_community.embeddings import HuggingFaceInferenceAPIEmbeddings

embeddings = HuggingFaceInferenceAPIEmbeddings(
    api_url="http://bge-m3.ai-hub.svc:8080/embed",
    model_name="BAAI/bge-m3"
)

# Generate embeddings
vectors = embeddings.embed_documents(["Hello", "World"])

---

Hybrid Retrieval

# Dense + Sparse search in Milvus
from pymilvus import AnnSearchRequest, WeightedRanker

# Get embeddings
response = requests.post(
    "http://bge-m3.ai-hub.svc:8080/embed",
    json={"texts": [query], "return_sparse": True}
)
dense_vec = response.json()["dense"][0]
sparse_vec = response.json()["sparse"][0]

# Parallel search
dense_req = AnnSearchRequest(dense_vec, "dense_vector", {"metric_type": "COSINE"}, 20)
sparse_req = AnnSearchRequest(sparse_vec, "sparse_vector", {"metric_type": "IP"}, 20)

# Combine results
results = collection.hybrid_search(
    [dense_req, sparse_req],
    rerank=WeightedRanker(0.7, 0.3),
    limit=10
)

---

Performance Tuning

Parameter	Description	Default
max_batch_size	Max texts per request	32
max_length	Max token length	8192
normalize	L2 normalize vectors	true

---

Monitoring

Metric	Query
Embed latency	bge_embed_duration_seconds
Rerank latency	bge_rerank_duration_seconds
Batch size	bge_batch_size
GPU memory	nvidia_gpu_memory_used_bytes

---

Consequences

Positive:

• State-of-the-art embeddings
• Multilingual support (Arabic)
• Hybrid dense + sparse
• Cross-encoder reranking
• Self-hosted

Negative:

• GPU required for performance
• Memory-intensive
• Batch size limits

---

Part of OpenOva