Changelog¶

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

2.0.0 - 2026-02-20¶

Breaking: Replaced port-based stdio JSON communication with erlang_python 1.5.0 NIF integration
All Python providers now use direct py:call instead of subprocess communication
2.7x improvement in batch throughput (936 vs 348 texts/sec with bge-small-en-v1.5)
Requires erlang_python 1.5.0+ as a dependency

If you were using custom provider configurations, the API remains the same. The venv option is still supported and recommended:

{ok, State} = barrel_embed:init(#{
    embedder => {local, #{
        venv => "/path/to/.venv"
    }}
}).

Added venv configuration option for all Python providers (local, fastembed, splade, colbert, clip)
Proper venv activation in port environment (sets VIRTUAL_ENV, PATH, PYTHONPATH)
New scripts/setup_venv.sh for fast venv setup using uv
Requirements files for different installation profiles:
priv/requirements.txt - Default (sentence-transformers + uvloop)
priv/requirements-minimal.txt - Minimal (no ML libs)
priv/requirements-full.txt - All providers
Documentation: docs/venv-setup.md

setup_python_venv.sh now uses uv when available (falls back to pip)
Python queue default limit changed from schedulers/2 + 1 to schedulers * 2 + 1
Updated all Python provider documentation with venv examples

cohere - Cohere Embed API with input type optimization
voyage - Voyage AI for RAG and domain-specific embeddings (code, law, finance)
jina - Jina AI with 8K context and free tier
mistral - Mistral AI with EU data residency
azure - Azure OpenAI for enterprise compliance
bedrock - AWS Bedrock (Titan, Cohere models) with IAM and API key auth
vertex - Google Vertex AI for GCP ecosystem

Updated hackney dependency to 2.0.1 for HTTP/2 support
Provider init now properly loads modules before checking exports
Removed redundant application:ensure_all_started(hackney) from providers (hackney starts via app.src)

Initial release extracted from barrel_vectordb
Core embedding coordinator (barrel_embed) with provider chain and fallback support
Provider behaviour (barrel_embed_provider) for implementing custom providers
Python execution rate limiter (barrel_embed_python_queue)

local - Local Python with sentence-transformers
ollama - Ollama server API (supports both /api/embed and /api/embeddings)
openai - OpenAI Embeddings API
fastembed - FastEmbed ONNX-based embeddings (lighter than sentence-transformers)
splade - SPLADE sparse embeddings for hybrid search
- embed_sparse/2, embed_batch_sparse/2 for native sparse vectors
- Automatic sparse-to-dense conversion for compatibility
colbert - ColBERT multi-vector embeddings for fine-grained matching
- embed_multi/2, embed_batch_multi/2 for token-level vectors
- maxsim_score/2 for late interaction scoring
clip - CLIP image/text cross-modal embeddings
- embed_image/2, embed_image_batch/2 for image embeddings
- Text embeddings in same vector space for cross-modal search