Director-AI
Response-level LLM Hallucination Guardrail
Score every claim in your LLM output against your knowledge base, with calibrated confidence and auditable evidence. NLI + RAG fact-checking, explicit abstention, prompt-injection detection, sealed evidence packets, and bounded streaming contradiction checks across a 530-module capability surface.
75.8%
Balanced accuracy
14.6 ms
Per claim (GPU)
782
Test files
9.4×
Rust speedup
12
SDK integrations
The problem
LLMs hallucinate. Your users trust them anyway. One wrong medical dosage. One fabricated legal citation. One invented financial figure. By the time a human reviewer catches it, the damage is done. Generic output filters catch obvious toxicity but miss subtle factual errors — the kind that sound perfectly plausible. Director-AI scores every claim against your knowledge base, returns an auditable verdict before the output is trusted, and — opt-in — halts streamed claims that contradict your grounding.
How it works
LLM Output
→
Claim Extraction
→
NLI Scoring
FactCG 0.4B
→
FactCG 0.4B
RAG Fact-Check
Your knowledge base
→
Your knowledge base
Dual Entropy
Confidence + divergence
→
Confidence + divergence
■ Halt stream
/
✓ Pass
Core features
Opt-in streaming contradiction check
Halts streamed claims that contradict your retrieved grounding facts. Response-level scoring is the production gate; the streaming check is opt-in and evidence-bound — not a sole guarantee.
Dual-entropy scoring
NLI contradiction detection (FactCG-DeBERTa, 0.4B params) combined with RAG fact-checking against your knowledge base. Two independent signals, one confidence score.
Injection detection
Intent-grounded, two-stage prompt injection detection: fast regex pre-filter + bidirectional NLI semantic analysis. 25 adversarial attack patterns tested.
Structured output verification
JSON schema validation, numeric consistency checking, reasoning chain verification, temporal freshness scoring. All stdlib-only, zero dependencies.
20 Rust accelerators
Performance-critical functions compiled to native via backfire-kernel (PyO3 FFI). Sanitiser 27×, temporal freshness 21×, confidence scoring 33× faster than pure Python.
EU AI Act compliance
Audit trails, adversarial robustness testing, domain presets (medical/finance/legal/creative), drift detection, and feedback loops. Built for regulated industries.
Integrations
Drop-in guards for every major LLM provider and framework. Zero code changes with the REST proxy.
LLM providers (SDK guards)
OpenAIAnthropic (Claude)AWS BedrockGoogle GeminiCohere
Frameworks
LangChainLlamaIndexLangGraphHaystackCrewAIDSPySemantic Kernel
Deployment
FastAPI middlewareREST/gRPC proxyDocker (CPU/GPU)Kubernetes HelmVoice AI (ElevenLabs/Deepgram)
Benchmarks
Accuracy (LLM-AggreFact, 29,320 samples)
| Scorer | Params | Balanced Accuracy | Latency (GPU) |
|---|---|---|---|
| FactCG-DeBERTa | 0.4B | 75.8% | 14.6 ms/pair |
| MiniCheck-Flan-T5-L | 0.8B | 77.4% | ~40 ms/pair |
| Heuristic-only (no NLI) | 0 | ~55% | <0.5 ms |
Latency (p99, 16-pair batch)
| Hardware | Backend | Latency |
|---|---|---|
| NVIDIA GTX 1060 | ONNX CUDA | 17.9 ms/pair |
| AMD RX 6600 XT | ROCm | 80.1 ms/pair |
| AMD EPYC 9575F | CPU | 118.9 ms/pair |
| Intel Xeon E5-2640 | CPU | 207.3 ms/pair |
Rust acceleration (backfire-kernel, 5000 iterations)
| Function | Python | Rust | Speedup |
|---|---|---|---|
| sanitiser_score | 57 µs | 2.1 µs | 27× |
| probs_to_confidence | 486 µs | 15 µs | 33× |
| temporal_freshness | 53 µs | 2.5 µs | 21× |
| lite_score | 47 µs | 26 µs | 1.8× |
| Geometric mean (12 functions) | 9.4× | ||
Quick start
# Install
pip install director-ai[all]
# Score a claim against a source
from director_ai import score
result = score("The Earth is 4.5 billion years old", "The Earth formed approximately 4.54 billion years ago.")
print(result) # GuardResult(score=0.94, passed=True)
# Or run as a REST proxy (zero code changes to your app)
director-ai serve --port 8000 --upstream https://api.openai.com/v1
NLI models
FactCG-DeBERTa-v3-Large
Default scorer. 0.4B params, MIT licensed. Best speed/accuracy trade-off. ONNX + TensorRT GPU acceleration paths available.
MiniCheck-Flan-T5-L
0.8B params. Higher accuracy (77.4%) at ~3× latency cost. Best for offline batch verification.
MiniCheck-DeBERTa-L
0.4B params. Alternative DeBERTa backbone with different NLI training data.
Gemma 4 E4B (LLM-as-judge)
LLM-based scoring for complex claims. Highest accuracy but sends data to external provider. Off by default.
Heuristic-only (Lite)
Zero-dependency scorer using word overlap, numeric consistency, and structural checks. <0.5 ms. ~55% accuracy. CPU-only fallback.
Rust backend (backfire)
Native compiled compute via backfire-kernel. 12 accelerated functions. No Python GIL. No CUDA dependency for basic scoring.
Domain presets
Medical
Strict thresholds. Dosage verification. Citation requirements. HIPAA-aware logging.
Finance
Numeric precision. Temporal freshness. Market data validation. FINMA-compatible audit trails.
Legal
Citation verification. Precedent checking. Jurisdiction awareness. Privilege-safe logging.
Creative
Relaxed thresholds. Factual claims still checked but creative expression permitted.
Licensing
Apache-2.0 Core
Free
The public core is free for commercial, closed-source, research, and personal use under Apache-2.0.
- Guardrail engine and five-tier scoring
- RAG grounding and evidence packets
- REST, gRPC, SDK, and framework integrations
- Rust-accelerated core paths
- Community support
- No core licence key or telemetry
BUSL-1.1 Advanced & Labs
From CHF 490/year
Pro and Full plans add private delivery and production entitlement for the source-available Advanced, Labs, and Research capability sets.
- Advanced production entitlement
- Full plan includes Labs and Research
- Private package-index delivery
- Priority support
- Custom model fine-tuning
- On-premise deployment assistance
Architecture at a glance
530 Python capability modules
1,057 capability classes and 227 public API exports in the v3.21.0 generated inventory.
83 Rust PyO3 bindings
Native compute surfaces with explicit package and licence boundaries.
17+ CLI commands
serve, proxy, bench, tune, finetune, batch, review, adversarial-test, doctor...
782 test files
Static release inventory; performance and accuracy claims remain tied to their separate evidence artefacts.
Python ≥3.11
Tested on 3.11, 3.12, 3.13. Zero core dependencies (numpy + requests only).
62 optional extras
Install only what you need: NLI, vector databases, server, SDKs, voice, enterprise, ONNX, and specialised integrations.
Director-AI Studio
Present and inspect streaming-AI safety examples, with a path from an educational demonstration to an evidence review.
Phase 0 scaffold · integration under development
See capabilities, a typical workflow and access · Explore all Studios
Choose your first step
Evaluate a model response against evidence and inspect the score, decision and abstention.
Read the installation guide or example · Discuss an integration
What you need
Use a Python environment and the dependencies specified by the installation guide. Supply your own response and relevant evidence.
How to interpret the result
A review decision is specific to its evidence and calibration. It does not guarantee that an arbitrary model response is true.
Follow this topic by RSS or confirmed email subscription. Keep a personal reading list.