v1.2.0

What's New

Score Caching

LRU cache with blake2b key hashing and configurable TTL — eliminates redundant NLI/embedding computation during streaming.

scorer = CoherenceScorer(threshold=0.6, cache_size=1024, cache_ttl=300)

Framework Integrations

• LangGraph — director_ai_node() factory + director_ai_conditional_edge() for state graph routing
• Haystack 2.x — DirectorAIChecker component with run(query, replies) interface
• CrewAI — DirectorAITool for agent-based workflows

Quantized NLI

8-bit model loading via bitsandbytes for <80ms GPU inference:

pip install director-ai[quantize]

scorer = CoherenceScorer(nli_quantize_8bit=True, nli_device="cuda")

Upgraded Embeddings

SentenceTransformerBackend with BAAI/bge-large-en-v1.5 (default) for higher-quality vector retrieval:

pip install director-ai[embeddings]

MkDocs Documentation Site

Full documentation at anulum.github.io/director-ai:

• User guides (scoring, streaming, evidence, config)
• 6 integration pages (SDK, LangChain, LlamaIndex, LangGraph, Haystack, CrewAI)
• API reference (mkdocstrings)
• Production deployment guides (scaling, Docker, Prometheus)
• Domain cookbooks (legal, medical, finance)

Enhanced Demo

Side-by-side comparison tab with token-level HTML highlighting (green/yellow/red + strikethrough for halted tokens).

Community

• GitHub issue templates (bug report, feature request)
• PR template
• Good first issues — 10 starter tasks for contributors

Stats

• 374 tests passing across Python 3.10/3.11/3.12
• 50 files changed, 2,464 insertions
• CI: lint, format, mypy, security audit — all green

Install

pip install director-ai==1.2.0

Full Changelog: v1.1.0...v1.2.0

v1.1.0

What's New

Pluggable NLI Backends

• NLIScorer(backend="minicheck") — MiniCheck-DeBERTa-L as alternative to default DeBERTa
• Graceful fallback to heuristic when backend package not installed

Native LLM Providers

• CoherenceAgent(provider="openai") — reads OPENAI_API_KEY from env
• CoherenceAgent(provider="anthropic") — reads ANTHROPIC_API_KEY from env
• Backward-compatible: llm_api_url= and default mock still work

CLI Benchmark Runner

• director-ai eval --dataset aggrefact --max-samples 100 --output results.json
• Delegates to benchmark suite with comparison table output

Streaming Halt Callbacks

• StreamingKernel(on_halt=callback) — fires with StreamSession on halt
• SafetyKernel(on_halt=callback) — fires with score on halt

SQLite Usage Dashboard

• GET /v1/stats — summary statistics
• GET /v1/stats/hourly — hourly breakdown
• GET /v1/dashboard — inline HTML dashboard
• Auto-records all reviews via the API server

Fixes

• CI type check now excludes benchmarks directory from mypy
• Docs workflow no longer requires deleted [research] extra

Full Changelog: v1.0.0...v1.1.0

v1.0.0 — Production Stable Release

v1.0.0 — Production Stable Release

Director-Class AI is a real-time LLM hallucination guardrail.

Highlights

• Enterprise modules: Policy, AuditLogger, TenantRouter, InputSanitizer
• Async scorer: CoherenceScorer.areview()
• LangChain + LlamaIndex integrations
• Optional extras: [langchain], [llamaindex], [server], [train], [nli], [vector]
• CLI: director-ai serve | score | batch | ingest
• Backfire Kernel: 6-crate Rust workspace (PyO3 FFI, UPDE, SEC, L16, SSGF, observers)

Research Purge (complete)

All research modules permanently removed from the package:

• src/director_ai/research/ deleted
• core/bridge.py deleted
• [research] extra deleted
• All research docs, notebooks, and Rust "consciousness" naming purged
• Proof: find . -name "*research*" -o -name "*consciousness*" returns zero results

Test Results

• Python: 290 passed, 8 skipped
• Rust: 152 passed

pip install director-ai

v0.9.0

v0.9.0

Added

• DeBERTa fine-tuning pipeline — data pipeline, training scripts, GPU/chunked scoring
• AggreFact benchmark suite — 5 result files (baseline, fine-tuned base/large, comparison)
• Core hardening — 24 test-enforced items: lazy torch import, thread safety, _clamp utility, singular metric guard, 1-D eigvecs fallback, history cap, NLI RuntimeError, etc.
• Test count: 144 → 375 (2.6x increase)

Changed

• CI: switched formatter from black to ruff format
• Backfire-kernel FFI refactored, slop cleanup

Fixed

• NLI data pipeline: ClassLabel cast for stratified split
• NLI scorer: None guards for mypy attr-defined errors
• 24 hardening test failures resolved

v0.8.2

Real benchmark results, code quality cleanup

v0.8.1

What's Changed

Added

• Semantic off-topic gate: _semantic_divergence() in CoherenceScorer uses sentence-transformers cosine similarity to detect off-topic responses before falling through to string-match for value accuracy
• Directory ingestion: VectorGroundTruthStore.ingest_from_directory(path, glob) recursively reads .txt, .md, and .jsonl files

Changed

• ChromaBackend now requires sentence-transformers (raises ImportError instead of silent fallback)
• Factual scoring cascade: NLI → semantic gate (off-topic detection) → string match
• README: removed pending benchmark placeholders, condensed research section

Details

• 397 tests passing, ruff clean
• Full changelog: CHANGELOG.md

v0.8.0 — Credibility Release

Highlights

• Configurable GroundTruthStore — no more hardcoded demo facts; pass facts= to constructor or use SAMPLE_FACTS for the original data
• Real semantic retrieval — ChromaBackend now uses SentenceTransformerEmbeddingFunction (all-MiniLM-L6-v2) with graceful fallback
• Streaming coherence fix — process_streaming() evaluates accumulated text, not individual tokens
• Bulk ingestion — VectorGroundTruthStore.ingest(texts, metadatas) + CLI director-ai ingest command (.txt / .jsonl)
• Backfire Kernel — Rust safety gate: 6,429 LOC, 153 tests, 29 Criterion benchmarks, 7,609x gradient speedup via analytic Jacobian
• LangChain integration — CoherenceCallbackHandler (optional [langchain] extra)
• Consumer-focused README — "drop-in coherence guardrail" positioning; SCPN research in collapsible section

Breaking Changes

• GroundTruthStore() now starts empty. Use GroundTruthStore(facts=SAMPLE_FACTS) for the old behaviour.

Install

pip install director-ai==0.8.0

Stats

• 397 tests passing (Python 3.10–3.12)
• ruff + black + mypy clean
• ~6,429 Rust LOC in backfire-kernel (not included in PyPI wheel — separate release)

Full changelog: https://github.com/anulum/director-ai/blob/v0.8.0/CHANGELOG.md