Vector similarity search platform for BharatML Stack.
Skye enables fast semantic retrieval by representing data as vectors and querying nearest matches in high-dimensional space. It is composed of three runnable components: skye-admin, skye-consumers, and skye-serving.
- 🔌 Pluggable Vector Databases — Support for multiple vector DB backends (Qdrant, NGT, Eigenix) via a generic abstraction layer
- 🏗️ Shared Embeddings, Isolated Indexes — Models are stored once but serve multiple tenants (variants), reducing data redundancy
- ⚡ Event-Driven Administration — Model lifecycle management through Kafka-based event flows for resilience and fault tolerance
- 💾 Multi-Layer Caching — In-memory (FreeCache) + distributed (Redis) caching for ultra-low-latency serving
- 🔍 Similarity Search APIs — gRPC APIs for similar-candidate search, bulk embedding retrieval, and dot-product computation
- 🔄 Real-Time + Batch Ingestion — Kafka consumers for both reset/delta batch jobs and real-time embedding updates
- 🎯 Configurable Distance Functions — DOT, Cosine, and Euclidean distance support
- 🛡️ Resilience — Circuit breakers, retry topics, and snapshot-based recovery
Skye is built around three components:
| Component | Role |
|---|---|
| skye-serving | Handles real-time similarity search queries with in-memory caching and vector DB lookups (gRPC, port 9090) |
| skye-consumers | Processes embedding ingestion (reset/delta jobs) and real-time aggregation events from Kafka (HTTP, port 8080) |
| skye-admin | Manages model lifecycle, onboarding, variant registration, and coordinates jobs (HTTP, port 8080) |
For detailed architecture and data flow diagrams, see the Skye documentation.
| Dependency | Purpose |
|---|---|
| etcd | Dynamic model/variant configuration |
| Kafka | Embedding ingestion events, model state machine |
| Qdrant | Vector database (pluggable) |
| ScyllaDB | Embedding storage + aggregator data |
| Redis | Distributed caching |
| RPC | Description |
|---|---|
GetSimilarCandidates(SkyeRequest) | Find similar candidates using embeddings or candidate IDs |
| RPC | Description |
|---|---|
GetEmbeddingsForCandidates(SkyeBulkEmbeddingRequest) | Bulk embedding retrieval for candidate IDs |
GetDotProductOfCandidatesForEmbedding(EmbeddingDotProductRequest) | Compute dot products between an embedding and candidates |
| Endpoint | Description |
|---|---|
GET /health | Health check |
| Endpoint | Description |
|---|---|
POST /api/v1/model/register-model | Register a new model |
POST /api/v1/model/register-variant | Register a variant for a model |
POST /api/v1/model/register-store | Register a storage store |
POST /api/v1/model/register-frequency | Register job frequency |
POST /api/v1/model/register-entity | Register an entity type |
| Endpoint | Description |
|---|---|
POST /api/v1/qdrant/create-collection | Create a Qdrant collection |
POST /api/v1/qdrant/process-model | Process a model (reset) |
POST /api/v1/qdrant/process-multi-variant | Process multiple variants |
POST /api/v1/qdrant/promote-variant | Promote variant to scale-up cluster |
POST /api/v1/qdrant/trigger-indexing | Trigger indexing pipeline |
- Go SDK — Client library for backend services
- Go 1.24 or later
- Docker and Docker Compose (for local development)
librdkafka-dev(CGO dependency for Kafka client)- A running BharatML Stack environment (etcd, Kafka, Qdrant, ScyllaDB, Redis)
The easiest way to run Skye with all its dependencies is via the BharatML Stack quick-start:
cd quick-start
./start.shThis starts Skye alongside all required services. See the Quick Start Guide for details.
cd skye
# Build all components
go build -o bin/skye-admin ./cmd/admin
go build -o bin/skye-consumers ./cmd/consumers
go build -o bin/skye-serving ./cmd/serving
# Run (example: serving)
./bin/skye-servingSkye is configured via environment variables (loaded through Viper). Dynamic model/variant configuration is managed via etcd.
| Variable | Description |
|---|---|
app_name | Application name |
app_env | Environment (staging/production) |
port | HTTP/gRPC server port |
auth_tokens | Authentication tokens |
| Variable | Description |
|---|---|
etcd_server | etcd server address |
etcd_username | etcd username |
etcd_password | etcd password |
etcd_watcher_enabled | Enable config hot-reload |
| Variable | Description |
|---|---|
kafka_broker | Kafka broker address |
kafka_group_id | Kafka consumer group ID |
kafka_topic | Kafka topic |
embedding_consumer_kafka_ids | Comma-separated embedding consumer IDs |
realtime_consumer_kafka_ids | Comma-separated real-time consumer IDs |
realtime_producer_kafka_id | Real-time producer ID |
| Variable | Description |
|---|---|
redis_addr | Redis server address |
redis_password | Redis password |
redis_db | Redis database number |
| Variable | Description |
|---|---|
storage_aggregator_db_count | Number of aggregator database connections |
storage_embedding_store_count | Number of embedding store connections |
cd skye
# Build images
docker build -f cmd/admin/Dockerfile -t skye-admin:latest .
docker build -f cmd/consumers/Dockerfile -t skye-consumers:latest .
docker build -f cmd/serving/Dockerfile -t skye-serving:latest .# Run (example: serving)
docker run -p 9090:9090 --env-file .env skye-serving:latest
# Run (example: admin)
docker run -p 8080:8080 --env-file .env skye-admin:latest| Version | Link |
|---|---|
| v1.0.0 | Skye Documentation |
Contributions are welcome! Please check our Contribution Guide for details on how to get started.
We encourage you to:
- Join our Discord community to discuss features, ideas, and questions
- Check existing issues before opening a new one
- Follow our coding guidelines and pull request process
- Participate in code reviews and discussions
- 💬 Discord: Join our community chat
- 🐛 Issues: Report bugs and request features on GitHub Issues
- 📧 Email: Contact us at ml-oss@meesho.com
BharatMLStack is open-source software licensed under the BharatMLStack Business Source License 1.1.
Built with ❤️ for the ML community from Meesho
If you find this useful, ⭐️ the repo — your support means the world to us!