director-ai 2.3.0

Real-time LLM hallucination guardrail — NLI + RAG fact-checking with token-level streaming halt

Director-AI

Real-time LLM hallucination guardrail — NLI + RAG fact-checking with token-level streaming halt

---

What It Does

Director-AI sits between your LLM and the user. It scores every output for hallucination before it reaches anyone — and can halt generation mid-stream if coherence drops below threshold.

graph LR
    LLM["LLM<br/>(any provider)"] --> D["Director-AI"]
    D --> S["Scorer<br/>NLI + RAG"]
    D --> K["StreamingKernel<br/>token-level halt"]
    S --> V{Approved?}
    K --> V
    V -->|Yes| U["User"]
    V -->|No| H["HALT + evidence"]

Three things make it different:

1. Token-level streaming halt — not post-hoc review. Severs output the moment coherence degrades.
2. Dual-entropy scoring — NLI contradiction detection (DeBERTa) + RAG fact-checking against your knowledge base.
3. Your data, your rules — ingest your own documents. The scorer checks against your ground truth.

Scope

100% Python — no compiled extensions required. Works on any platform with Python 3.10+.

Layer	Packages	Install
Core (zero heavy deps)	CoherenceScorer, StreamingKernel, GroundTruthStore, SafetyKernel	pip install director-ai
NLI models	DeBERTa, FactCG, MiniCheck, ONNX Runtime	pip install director-ai[nli]
Vector DBs	ChromaDB, Pinecone, Weaviate, Qdrant	pip install director-ai[vector]
LLM judge	OpenAI, Anthropic escalation	pip install director-ai[openai]
Observability	OpenTelemetry spans	pip install director-ai[otel]
Server	FastAPI + Uvicorn	pip install director-ai[server]

Quickstart

Method	Command
pip install	pip install director-ai
CLI scaffold	director-ai quickstart --profile medical
Colab
HF Spaces	Try it live
Docker	docker run -p 8080:8080 ghcr.io/anulum/director-ai:latest

6-line guard

from director_ai import guard
from openai import OpenAI

client = guard(
    OpenAI(),
    facts={"refund_policy": "Refunds within 30 days only"},
    threshold=0.6,
)

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "What is the refund policy?"}],
)

Catch and inspect a halt

from director_ai import guard, HallucinationError
from openai import OpenAI

client = guard(OpenAI(), facts={"policy": "Refunds within 30 days only"})

try:
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": "What is the refund policy?"}],
    )
except HallucinationError as exc:
    print(f"HALTED: coherence={exc.score.score:.3f}")
    print(f"Evidence: {exc.score.evidence}")

Score a response

from director_ai.core import CoherenceScorer, GroundTruthStore

store = GroundTruthStore()
store.add("sky color", "The sky is blue due to Rayleigh scattering.")

scorer = CoherenceScorer(threshold=0.6, ground_truth_store=store)
approved, score = scorer.review("What color is the sky?", "The sky is green.")

print(approved)     # False
print(score.score)  # 0.42

Streaming halt

from director_ai.core import StreamingKernel

kernel = StreamingKernel(hard_limit=0.4, window_size=5)
session = kernel.stream_tokens(token_generator, lambda tok: my_scorer(tok))

if session.halted:
    print(f"Halted at token {session.halt_index}: {session.halt_reason}")

Installation

pip install director-ai                      # heuristic scoring
pip install director-ai[nli]                 # NLI model (DeBERTa)
pip install director-ai[vector]              # ChromaDB knowledge base
pip install "director-ai[nli,vector,server]" # production stack

Framework integrations: [langchain], [llamaindex], [langgraph], [haystack], [crewai].

Full installation guide: docs.

Docker

docker run -p 8080:8080 ghcr.io/anulum/director-ai:latest        # CPU
docker run --gpus all -p 8080:8080 ghcr.io/anulum/director-ai:gpu # GPU

Benchmarks

Accuracy — LLM-AggreFact (29,320 samples)

Model	Balanced Acc	Params	Latency	Streaming
Bespoke-MiniCheck-7B	77.4%	7B	~100 ms	No
Director-AI (FactCG)	75.8%	0.4B	14.6 ms	Yes
MiniCheck-Flan-T5-L	75.0%	0.8B	~120 ms	No
MiniCheck-DeBERTa-L	72.6%	0.4B	~120 ms	No

75.8% balanced accuracy at 17x fewer params than the leader. 14.6 ms/pair with ONNX GPU batching — faster than every competitor at this accuracy tier. Director-AI's unique value is the system: NLI + KB + streaming halt.

Full results: benchmarks/comparison/COMPETITOR_COMPARISON.md.

Domain Presets

8 built-in profiles with tuned thresholds:

director-ai config --profile medical   # threshold=0.75, NLI on, reranker on
director-ai config --profile finance   # threshold=0.70, w_fact=0.6
director-ai config --profile legal     # threshold=0.68, w_logic=0.6
director-ai config --profile creative  # threshold=0.40, permissive

Known Limitations

1. Heuristic fallback is weak: Without [nli], scoring uses word-overlap heuristics (~55% accuracy). Use strict_mode=True to reject (0.9) instead of guessing.
2. Summarisation is a weak spot: NLI models under-perform on summarisation (AggreFact-CNN: 68.8%, ExpertQA: 59.1%).
3. ONNX CPU is slow: 383 ms/pair without GPU. Use onnxruntime-gpu for production.
4. Weights are domain-dependent: Default w_logic=0.6, w_fact=0.4 suits general QA. Adjust for your domain.
5. Chunked NLI: Very short chunks (<3 sentences) may lose context.
6. LLM-as-judge sends data externally: When llm_judge_enabled=True, truncated prompt+response (500 chars) are sent to the configured provider (OpenAI/Anthropic). Do not enable in privacy-sensitive deployments without user consent.

Migrating from 1.x

1.x name	2.x name	Notes
DirectorModule	CoherenceScorer	Same API, new name
BackfireKernel	SafetyKernel	Same API, new name
StrangeLoopAgent	CoherenceAgent	Same API, new name
KnowledgeBase	GroundTruthStore	Same API, new name
MockActor	MockGenerator	Same API, new name
RealActor	LLMGenerator	Same API, new name

Old names still work but emit DeprecationWarning. They will be removed in 3.0.

Breaking changes in 2.3.0:

• strict_mode=True now rejects (divergence=0.9) when NLI is unavailable, instead of returning neutral 0.5.
• guard() uses duck-type detection instead of module-name checks. Custom clients that expose client.chat.completions.create are now accepted.
• Enterprise modules are lazy-loaded since 2.2.0 — import director_ai no longer pulls heavy deps.

Citation

@software{sotek2026director,
  author    = {Sotek, Miroslav},
  title     = {Director-AI: Real-time LLM Hallucination Guardrail},
  year      = {2026},
  url       = {https://github.com/anulum/director-ai},
  version   = {2.3.0},
  license   = {AGPL-3.0-or-later}
}

License

Dual-licensed:

1. Open-Source: GNU AGPL v3.0 — research, personal use, open-source projects.
2. Commercial: Proprietary license — removes copyleft for closed-source and SaaS.

See Licensing for pricing tiers and FAQ.

Contact: anulum.li/contact | [e-mail]

Contributing

See CONTRIBUTING.md. By contributing, you agree to AGPL v3 terms.