Skip to content

Installation

Base Package

pip install director-ai

Includes: coherence scorer, streaming kernel, safety kernel, ground truth store, heuristic scoring.

Optional Extras

Extra Command Adds
nli pip install director-ai[nli] FactCG-DeBERTa-v3-Large NLI (75.8% bal. acc, 14.6ms/pair ONNX GPU)
minicheck pip install director-ai[minicheck] MiniCheck alternative (72.6% bal. acc)
vector pip install director-ai[vector] ChromaDB vector store
embeddings pip install director-ai[embeddings] sentence-transformers (bge-large)
onnx pip install director-ai[onnx] ONNX Runtime inference (14.6ms/pair GPU, portable CPU fallback)
quantize pip install director-ai[quantize] bitsandbytes 8-bit quantization
openai pip install director-ai[openai] OpenAI SDK guard
anthropic pip install director-ai[anthropic] Anthropic SDK guard
langchain pip install director-ai[langchain] LangChain integration
llamaindex pip install director-ai[llamaindex] LlamaIndex integration
langgraph pip install director-ai[langgraph] LangGraph integration
haystack pip install director-ai[haystack] Haystack integration
crewai pip install director-ai[crewai] CrewAI integration
finetune pip install director-ai[finetune] Domain-specific NLI fine-tuning (torch, transformers, datasets)
grpc pip install director-ai[grpc] gRPC server (grpcio, protobuf)
server pip install director-ai[server] FastAPI server
otel pip install director-ai[otel] OpenTelemetry tracing
enterprise pip install director-ai[enterprise] Multi-tenant, audit, Redis
docs pip install director-ai[docs] MkDocs documentation tools
dev pip install director-ai[dev] pytest, ruff, mypy, sphinx 
pip install director-ai[nli,vector,embeddings,openai]

FactCG NLI (75.8% balanced accuracy, 14.6ms/pair ONNX GPU batch), ChromaDB + bge-large embeddings, OpenAI SDK interception.

GPU Acceleration

For GPU-accelerated NLI scoring:

pip install director-ai[nli,quantize]

Set device and dtype:

from director_ai import CoherenceScorer

scorer = CoherenceScorer(
    use_nli=True,
    nli_device="cuda",
    nli_torch_dtype="float16",
    nli_quantize_8bit=True,  # reduces VRAM, slightly slower than fp32
)

Python Version

Requires Python 3.11+.