Skip to content

Tutorials

Interactive Jupyter notebooks covering Director-AI from first principles to production deployment. Every notebook runs in Google Colab with zero local setup.

For a buyer- and use-case-oriented index across every published notebook, see the Notebook Gallery. For the product and commercial context behind the notebook tracks, start with Applications and Market Map.

What The Tutorial Set Covers

The notebooks are organised as a product learning path, not as isolated demos:

  • Foundations: scoring, evidence, and the coherence engine.
  • Application controls: SDK guard, opt-in streaming contradiction halt, domain presets, and structured verification.
  • RAG and domain applications: vector-backed facts, medical examples, and integration adapters.
  • Production and evaluation: batch processing, multi-tenant deployment, custom adaptation, and benchmark interpretation.

Use Evaluation Onboarding alongside the notebooks when you need pilot evidence rather than a feature tour.

Choose By Role

Role Best first notebook Why
First evaluator Quickstart Install, score, guard, and inspect halt metadata
Application developer Production Guardrails Wrap a real application path and choose failure modes
RAG engineer Vector RAG Pipeline Ground answers in private facts and retrieved evidence
Runtime engineer Streaming Halt Deep Dive Understand halt thresholds and token-by-token enforcement
Enterprise platform team Enterprise Multi-Tenant Inspect tenant, REST/gRPC, Docker, Kubernetes, and monitoring patterns
Evaluation engineer Batch Processing & Evaluation Build benchmark and regression-gate workflows

For a non-notebook pilot checklist, use Evaluation Onboarding.

Learning Outcomes

Track You should be able to prove
Foundations A governed fact can approve one response and reject a contradiction
Streaming safety A partial output can halt before the full answer is emitted
Retrieval A verdict can cite retrieved chunks instead of relying on prompt text alone
Integrations A real SDK or framework path can be wrapped without changing the model provider
Evaluation A labelled set can drive threshold and regression-gate decisions
Enterprise Tenant, auth, metrics, and deployment controls are visible before rollout

Tutorial-To-Application Map

Application Recommended tutorial path Pilot proof
Customer support Quickstart → Production Guardrails → Domain Presets Policy answer allowed, unsupported policy answer rejected
Enterprise RAG Vector Store → Vector RAG Pipeline → Batch Processing Retrieved evidence attached to pass/fail verdicts
Streaming assistant Streaming Oversight → Streaming Halt Deep Dive Completed contradictory claim halts before continuation
Regulated review Medical RAG Chatbot → Verification Gems → Evaluation Pipelines Human-review examples and threshold evidence
Platform deployment Production Guardrails → Enterprise Multi-Tenant Auth, tenant, audit, metrics, and rollback surfaces visible

Learning Path

graph TD
    subgraph "Foundations (30 min)"
        A["00 Quickstart<br/>5 min"] --> B["01 Coherence Engine<br/>15 min"]
    end

    subgraph "Core Tracks"
        B --> C["09 Production Guardrails<br/>guard() + SDK wrapping"]
        B --> D["02 Streaming Oversight<br/>claim-level halt"]
        B --> V["16 Verification Gems<br/>8 standalone modules"]
    end

    subgraph "Deep Dives"
        C --> E["10 Vector RAG Pipeline<br/>ChromaDB + reranking"]
        C --> F["12 Domain Presets & Config<br/>8 profiles"]
        D --> G["11 Streaming Halt Deep Dive<br/>hard/soft/trend"]
        E --> H["06 Medical RAG Chatbot<br/>domain-specific"]
    end

    subgraph "Production (enterprise)"
        F --> I["13 Batch Processing<br/>evaluation pipelines"]
        I --> J["14 Enterprise Multi-Tenant<br/>Docker + K8s"]
        J --> K["15 Custom Fine-Tuning<br/>domain NLI + ONNX export"]
    end

    style A fill:#7c4dff,color:#fff
    style B fill:#7c4dff,color:#fff
    style C fill:#2e7d32,color:#fff
    style D fill:#1565c0,color:#fff
    style V fill:#ff8f00,color:#fff
    style K fill:#c62828,color:#fff

Getting Started

Start here. These two notebooks teach the core concepts in under 30 minutes.

# Notebook What You Learn Time Colab
00 Quickstart Install, score, guard, stream, presets 5 min Colab
01 Coherence Engine CoherenceScorer, SafetyKernel, CoherenceAgent, dual-entropy formula 15 min Colab

Core Features

Deep dives into the four pillars of Director-AI.

# Notebook What You Learn Time Colab
09 Production Guardrails guard() for supported chat, message, cloud-runtime, Mistral, Cohere, and Pydantic AI, failure modes, streaming guards 20 min Colab
10 Vector RAG Pipeline Semantic retrieval, ChromaDB, pluggable backends, reranking, multi-tenant KB 25 min Colab
11 Streaming Halt Deep Dive Hard limit, sliding window, trend detection, async, per-token visualization 20 min Colab
12 Domain Presets & Config 8 profiles, env vars, YAML, backends, strict mode, multi-GPU, LLM-as-judge 15 min Colab

Domain Applications

Real-world integrations and domain-specific patterns.

# Notebook What You Learn Time Colab
02 Streaming Oversight StreamingKernel basics, token-by-token monitoring 10 min Colab
03 Vector Store VectorGroundTruthStore, InMemoryBackend, fact ingestion 10 min Colab
05 SSGF Geometry Self-similar geometry foundation concepts 10 min Colab
06 Medical RAG Chatbot Healthcare-specific guardrails, high thresholds, evidence citations 20 min Colab
07 LangChain Integration CoherenceCallbackHandler, chain integration, output parsing 15 min Colab
08 Provider Adapters Commercial SDK adapters, cloud-runtime adapters, agent frameworks, Guardrails AI, and Vercel AI SDK adapter patterns 10 min Colab

Verification & Analysis

Standalone analysis modules — no NLI model required. All stdlib-only.

# Resource What You Learn Time Colab
16 Verification Gems All 8 gems: numeric, reasoning, temporal, consensus, conformal, feedback loops, agentic, REST API 15 min Colab
— Guide: Verification Gems Full parameter reference for all 8 gems 15 min —
— Example: verification_gems_demo.py Runnable demo of all 7 standalone verification modules 5 min —

CLI quick start:

director-ai verify-numeric "Revenue grew 50% from \$100 to \$120"
director-ai verify-reasoning "Step 1: A is true. Step 2: Therefore B."
director-ai temporal-freshness "The CEO of Apple is Tim Cook"
director-ai check-step "Find revenue data" "search" "revenue Q3"

REST quick start (with server running):

curl -X POST http://localhost:8080/v1/verify/numeric \
  -H "Content-Type: application/json" \
  -d '{"text": "Revenue grew 50% from $100 to $120."}'

Enterprise & Production

Scale, evaluate, fine-tune, and deploy.

# Notebook What You Learn Time Colab
04 End-to-End Benchmark Full benchmark suite, latency profiling, accuracy metrics 15 min Colab
13 Batch Processing & Evaluation BatchProcessor, evaluation pipelines, claim attribution, regression gates 20 min Colab
14 Enterprise Multi-Tenant Tenant isolation, REST/gRPC servers, Docker, Kubernetes, monitoring 25 min Colab
15 Custom Fine-Tuning JSONL data prep, validation, training, anti-forgetting, ONNX export, REST API 30 min Colab

Prerequisites

All notebooks run on Python 3.11+ with pip install director-ai.

Notebooks requiring optional extras note this in their first cell:

Extra Install Notebooks
NLI scoring pip install director-ai[nli] 01, 04, 06, 09–15
Vector store pip install director-ai[vector] 10, 14
Fine-tuning pip install director-ai[finetune] 15
Server pip install director-ai[server] 14
gRPC pip install director-ai[grpc] 14

Running Locally

git clone https://github.com/anulum/director-ai.git
cd director-ai
pip install -e ".[dev,nli]"
jupyter lab notebooks/