CyborgDB Release Notes

​

CyborgDB v0.18.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript, TypeScript & Go)

​

New Features

• BM25 full-text and hybrid search
  • Mark metadata fields as full_text in metadata_schema to make them keyword-searchable; tune scoring per index with bm25_k1 / bm25_b
  • query(text=...) runs a keyword search, optionally limited to specific text_fields; combine it with a query vector for hybrid search, blended with alpha, rrf_k, and window_mult
  • Available in the embedded libraries, the REST API, and the Python, JavaScript/TypeScript, and Go SDKs
• Metadata schema. create_index accepts a metadata_schema that sets, per field, whether it is filterable, indexed for fast $regex / $contains matching (pattern), or full_text
• query_metadata returns the items matching a metadata filter without a query vector, in the embedded libraries, all client SDKs, and the REST API (POST /v1/vectors/query_metadata)
• TurboQuant storage precision. storage_precision adds the tq12, tq8, tq6, and tq4 tiers for much smaller stored vectors at a small recall cost (tq4 is the most compact)
• GET /v1/embedding-models lists the built-in embedding models the service supports
• load_sample_dataset (loadSampleDataset / LoadSampleDataset) in the client SDKs loads a sample dataset for quick evaluation
• StorageConfig.disk(...) adds a cache_metadata_index option

​

Improvements

• Built-in embedding models no longer require PyTorch; installing the [embeddings] extra is no longer needed
• $regex filters are now evaluated with RE2, giving linear-time matching and protection against pathological patterns
• $contains filter operator for substring matching
• The client SDKs now raise standardized error types (for example AuthenticationError, NotFoundError, ValidationError)
• Telemetry is not sent when no API key is set
• Index dimension is now validated at create_index

​

Breaking Changes

• Memory backend removed. StorageConfig.memory() and CYBORGDB_DB_TYPE=memory are gone; use disk (or s3) instead
• One process per storage target. A disk path or S3 bucket/prefix must be used by a single CyborgDB process at a time. Running multiple embedded processes or service replicas against the same storage target is not supported
• GET /v1/indexes/list and GET /v1/indexes/training-status now require the root (admin) key

​

Bug Fixes

• JavaScript/TypeScript SDK: fixed decoding of contents returned by get(), and verifySsl: false is now honored

​

CyborgDB v0.17.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript, TypeScript & Go)

​

New Features

• Access control (RBAC) on encrypted indexes, in both the embedded libraries and the service
  • The holder of the root index KEK can mint per-user keys scoped to read and/or write — the keys a user holds determine their permissions, enforced cryptographically on every request
  • Embedded: create_user_keys (Python) / CreateUserKeys (C++), plus index_key= / user_id= per-operation overrides for stateless deployments
  • Service: POST/GET/DELETE on /v1/indexes/{name}/users, surfaced as users in the Python, JavaScript/TypeScript, and Go SDKs
• Per-index KMS & BYOK. Each index can be wrapped by its own operator-managed key, optionally living in a customer’s own AWS account
  • Providers: aws-kms (AWS KMS Encrypt/Decrypt), aws (Secrets Manager), and none
  • Pass kms_name to create_index instead of index_key; the service resolves the key server-side from the index’s stored key envelope, so the SDK never holds a long-term key
  • Cross-account BYOK via role_arn + external_id; KMS credentials are kept separate from CYBORGDB_S3_* storage credentials
• Free tier. CYBORGDB_API_KEY is now optional — without it, CyborgDB runs free-tier with a 1,000,000-item-per-index cap. Demo keys are available via get_demo_api_key() / getDemoApiKey() / GetDemoAPIKey()
• YAML configuration. The service can be configured entirely by cyborgdb.yaml (resolved from CYBORGDB_CONFIG_FILE, the working directory, or /etc/cyborgdb/), with ${VAR} / ${VAR:-default} env-var substitution. Environment variables override the file
• storage_precision lets you store rerank vectors as float16 instead of float32, halving the on-disk footprint at a small recall cost
• rerank_mult on query tunes stage 1 retrieval depth (top_k * rerank_mult candidates before reranking), and include=["vector"] returns the stored rerank vector

​

Improvements

• n_lists is now selected automatically by the core engine

​

Breaking Changes

• Single index type. CyborgDB now ships a single IVF index type. The polymorphic index_config object is gone, along with the IVFFlat / IVFPQ / IVFSQ distinction
  • Configuration parameters (dimension, metric, embedding_model, storage_precision) are now top-level fields on create_index
  • dimension is optional — the server auto-detects it from the first upsert, or infers it from embedding_model
  • In the embedded SDKs, all parameters after index_key are keyword-only
• Storage refactor. The backing store surface is now just memory, disk, or s3
  • The standalone, postgres, and redis backends have been removed
  • CYBORGDB_CONNECTION_STRING is replaced by CYBORGDB_DISK_PATH and the CYBORGDB_S3_* variables — switch standalone → disk when upgrading
  • In the embedded SDKs, a single StorageConfig (StorageConfig.memory() / .disk(path) / .s3(bucket)) replaces the index_location / config_location / items_location trio of DBConfig objects
• Authentication overhaul. CYBORGDB_API_KEY is now a license key only and no longer gates access to the REST API
  • Request authentication is controlled by CYBORGDB_SERVICE_ROOT_KEY. Leave it unset and authentication is disabled; set it and routes accept only the root key or a per-user cdbk_… token in the X-API-Key header
  • The legacy cyborgdb-api-key header and the REQUIRE_API_KEY flag have been removed
• A wrong index_key now returns 401 instead of 500

​

CyborgDB v0.16.1

​

Behavior Changes

• query() now returns only id fields by default when include is not specified
  • Previously defaulted to ["distance", "metadata"]; new default is []
  • Pass include=["distance", "metadata"] explicitly to retain the prior behavior
  • Applies to both the embedded library and the REST service (QueryRequest, BatchQueryRequest, and BinaryQueryRequest all default include to [])
  • Breaking change for REST API clients: requests without an explicit include field will no longer receive distance or metadata in the response

​

Bug Fixes

• Fixed an S3 backing store bug to properly support S3 versioned objects
• Fixed a distance computation bug affecting query result ordering
• Fixed an arm64 Linux issue where cyborgdb-core could fail to load runtime dependencies
• query() now returns vector IDs by default

​

Improvements

• CUDA wheels now statically link CUDA runtime dependencies for improved portability

​

CyborgDB v0.16.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript, TypeScript & Go)

​

New Features

• CyborgDB now supports Amazon S3 as a backing store for encrypted indexes
  • Configure via CYBORGDB_DB_TYPE=s3 with an S3 connection string
  • Compatible with AWS S3, MinIO, and any S3-compatible object storage
• CyborgDB Migrate: New tool for migrating vector embeddings from other databases into CyborgDB
  • Supports Pinecone, Qdrant, Weaviate, ChromaDB, and Milvus as sources
  • Interactive TUI wizard or headless TOML-based config for CI/CD pipelines
  • Install via pip install cyborgdb-migrate
• IVFSQ default scalar quantization bits increased from 8 to 16 for improved recall accuracy

​

Improvements & Bug Fixes

• Stability improvements and bug fixes across all components
• Ingest errors now properly propagate to callers instead of being silently swallowed
• Improved serialization support for all backing store types
• New API contract test suite for guaranteed interface stability
• Expanded E2E test coverage across all SDK languages and backing stores

​

CyborgDB v0.15.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript, TypeScript & Go)

​

New Features

• Significant performance optimizations for all index types
• Added new IVFSQ index type and replace IVFFlat as the default
• Deprecated IVF index type
• Added standalone support for on-device encrypted indexes without external database dependencies
• Added support for binary upsert and query vectors for faster data transfer and reduced memory usage
• New /training-status endpoint in REST API for monitoring index training progress and status
• Improved error handling and logging across all components

​

CyborgDB v0.14.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript, TypeScript & Go)

​

New Features

• Significant performance optimizations for cyborgdb-core:
  • Reduced single-query (unbatched) latency by 75%
• Introduced standalone database mode for quick evaluation and testing:
  • Automatically enabled when no external database is configured
  • Useful for rapid prototyping and development without infrastructure overhead
  • Not recommended for production use - use PostgreSQL or Redis for production deployments
• Improved type definitions for TypeScript SDK

​

CyborgDB v0.13.1

​

Bug Fixes

• Fixed an issue where cyborgdb-core would fail to find runtime dependencies on certain Linux arm64 systems
• Exposed demo API key creation on Go and TypeScript SDKs

​

CyborgDB v0.13.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript, TypeScript & Go)

​

New Features

• Merged cyborgdb-lite functionality into cyborgdb-core:
  • cyborgdb-lite package has been deprecated and removed
  • All features previously available in cyborgdb-lite are now included in cyborgdb-core
  • Simplified package structure with a single unified library for embedded use
• CyborgDB Service now supports CUDA with pip install cyborgdb-service[cuda]:
  • Enables existing cyborgdb-core CUDA optimizations in the REST API microservice
• Deprecated support for Python 3.9 due to EOL for:
  • cyborgdb-core
  • cyborgdb-service
  • cyborgdb-py

​

CyborgDB v0.12.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript, TypeScript & Go)

​

New Features

• Go Client SDK now available:
  • go get github.com/cyborginc/cyborgdb-go
• Automatic index management enabled by default in CyborgDB Service:
  • Simplifies setup and reduces manual configuration
  • IndexConfig classes now optional - CyborgDB can auto-configure indexes on the fly based on data & usage characteristics
  • Training & retraining now handled automatically, optimizing index performance without manual intervention
  • Queries can now automatically determine the optimal number of clusters to search
• LangChain integration now available for Python Client SDK

​

CyborgDB v0.11.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker)
• Client SDK (Python, JavaScript & TypeScript)

​

New Features

• CyborgDB is now available as a REST API microservice
  • Available as a Python package (cyborgdb-service)
  • Available as a Docker image (cyborginc/cyborgdb-service)
• CyborgDB Client SDKs now available for REST API microservice
  • Python (cyborgdb)
  • JavaScript (cyborgdb)
  • TypeScript (cyborgdb)
• Introduced API key management and a new Admin Dashboard for enhanced operational control
• Optimized concurrency handling using opportunistic mutexes to improve performance under load
• Improved Docker container reliability and stability across deployment environments

​

CyborgDB v0.10.0

​

Deployment Model

• Embedded Library (Python & C++)
• REST API (Microservice via Python or Docker - early access)
• Client SDK (Python, JavaScript & TypeScript - early access)

​

New Features

• CyborgDB REST API service in early access
• CyborgDB Client SDKs in early access
  • Python
  • JavaScript
  • TypeScript
• LangChain integration for CyborgDB now available
  • Install via pip install cyborgdb-core[langchain]
  • Use via from cyborgdb_core.integrations.langchain import CyborgVectorStore
  • Supports both cyborgdb-core and cyborgdb-lite
• Added configurable logging utility for embedded library in C++ and Python

​

CyborgDB v0.9.2

​

New Features

• cyborgdb_lite is now available for evaluation and non-commercial use on PyPI

​

CyborgDB v0.9.1

​

Bug Fixes

• Fixed a bug where automatic embedding generation would not load the model correctly when using load_index()
• Fixed a bug which caused intermittent query() recall issues on Linux
• Fixed incorrect type returns for get() in C++ and Python APIs

​

Improvements

• Enabled pre-filtering on metadata queries to improve performance
• Added annotations and IDE auto-completion for the Python API
• Expanded support to Python 3.9 - 3.13

​

CyborgDB v0.9.0

​

New Features

• Added support for metadata fields & query filtering
  • Metadata fields can be added to items during upsert() calls
  • query() calls can filter results based on metadata fields
• Added support for item deletion
  • delete() calls can remove items from the encrypted index
  • upsert() conflicts are now resolved by overwriting existing items
• Added support for automatic embedding generation
  • An embedding model can be specified during index creation
  • upsert() calls can generate embeddings directly from index contents
  • query() calls can use the model to generate embeddings for query vectors

​

Breaking Changes

• IDs are now string-type (instead of int)
  • This impacts upsert(), get() and query() function signatures
• get_item() and get_items() have been replaced by get()

​

Cyborg Vector Search v0.8.0

​

New Features

• Added support for encrypted item storage & retrieval
  • Encrypted indexes can now handle item content storage
  • upsert() calls can take item content as bytes, encrypt and store them in the index
  • get_item() calls can retrieve encrypted item content and decrypt them
• Client-side caching now available for encrypted indexes (via max_cache_size parameter)
• Client can now handle more than one index via multiple EncryptedIndex objects
• Optimizations:
  • Full-pipeline GPU acceleration for encrypted indexes
  • Optimized quantization and ranking logic
  • Backing store batch operations for improved performance

​

Breaking Changes

• Python & C++ APIs now split into two classes:
  • Client class now handles DB backend connection, index creation and loading
  • EncryptedIndex class now handles data operations (e.g., upsert(), query(), etc.)

​

Cyborg Vector Search v0.7.0

​

New Features

• First public release of Cyborg Vector Search
• Python API available through PyBind11
• GPU acceleration with CUDA support