The only LLM guardrail that halts hallucinations mid-stream. NLI + RAG dual scoring against your own knowledge base. Enterprise-ready. Open-source.
Not post-hoc review. The safety kernel monitors coherence token-by-token and severs output the moment it degrades. No other guardrail does this.
NLI contradiction detection (DeBERTa) + RAG fact-checking against your own knowledge base. Both must pass. One model catching what the other misses.
Ingest PDFs, directories, or any text into a ChromaDB-backed knowledge base. The scorer checks LLM output against your ground truth, not a generic model.
| Feature | Director-AI | NeMo Guardrails | Guardrails AI | SelfCheckGPT |
|---|---|---|---|---|
| Streaming halt (token-level) | Yes | No | No | No |
| NLI contradiction detection | Yes | No | No | Yes |
| Custom knowledge base | Yes (ChromaDB) | Yes | Yes | No |
| LangChain integration | Built-in | Built-in | Built-in | Manual |
| LlamaIndex integration | Built-in | No | Built-in | No |
| Policy engine (YAML) | Yes | Colang | RAIL | No |
| Multi-tenant isolation | Yes | No | No | No |
| Audit trail | JSONL | No | No | No |
| Self-hosted / air-gapped | Yes | Yes | Yes | Yes |
| License | AGPL v3 + Commercial | Apache 2.0 | Apache 2.0 | MIT |
from director_ai.integrations.langchain import DirectorAIGuard
guard = DirectorAIGuard(
facts={"refund": "Refunds within 30 days."},
threshold=0.6,
)
chain = my_llm | guard # halts hallucinationsfrom director_ai.integrations.llamaindex import DirectorAIPostprocessor
pp = DirectorAIPostprocessor(
facts={"pricing": "Enterprise: $99/mo."},
threshold=0.6,
)
engine = index.as_query_engine(
node_postprocessors=[pp]
) # filters hallucinated nodesEvaluated on LLM-AggreFact (29,320 samples across 11 datasets)
| Tool | Bal. Acc | Params | Latency | Streaming |
|---|---|---|---|---|
| Bespoke-MiniCheck-7B | 77.4% | 7B | ~100 ms (GPU) | No |
| Director-AI (FactCG) | 75.8% | 0.4B | 575 ms (CPU) | Yes |
| MiniCheck-Flan-T5-L | 75.0% | 0.8B | ~120 ms | No |
| MiniCheck-DeBERTa-L | 72.6% | 0.4B | ~120 ms | No |
| HHEM-2.1-Open | 71.8% | ~0.4B | ~200 ms | No |
Honest assessment: 75.8% balanced accuracy ranks 4th on the LLM-AggreFact leaderboard — within 1.6pp of the top-scoring 7B model while using only 0.4B params. Director-AI's unique value is the system: NLI + KB facts + streaming token-level halt. No competitor offers real-time streaming gating. CPU latency (~575 ms with source chunking) drops to ~50-80 ms on GPU.
AGPL v3 open-source free forever. Commercial license removes copyleft.
Perpetual license: $1,299 one-time · First 50 licensees: 50% off first year
Install in 30 seconds. Score your first response in 60.