Release v2.0.0 — Enterprise Multimodal RAG Upgrade

🚀 Release v2.0.0 Enterprise Multimodal RAG Upgrade

This release marks the transition from a research prototype to a production‑ready enterprise system.

🔧 Architecture

• Refactored around pluggable Abstract Base Classes (BaseParser, BaseEmbedder, BaseVectorStore, etc.)
• Unified dependency injection for all components
• Centralized configuration via Pydantic v2 BaseSettings

📥 Ingestion

• PDF parsing with Unstructured.io, IBM Docling, and Marker
• Layout‑aware semantic chunker (tables + captions preserved)
• Vision fallback chain: GPT‑4o, Gemini 2.0 Flash, Qwen2‑VL, HuggingFace via vLLM
• PII redaction with Microsoft Presidio

🔍 Retrieval & Generation

• Hybrid retrieval: dense vectors + BM25 fused with RRF (k=60)
• Cross‑encoder reranker (Cohere Rerank v3)
• Program‑of‑Thought executor for financial arithmetic (AST‑validated, 5s timeout)
• Smart LLM routing: gpt‑4o‑mini vs gpt‑4o

🛡️ Guardrails & Security

• Numeric grounding: values must appear in retrieved context
• Prompt injection blocked at query time
• Immutable audit log with SHA‑256 tamper hashes
• Constant‑time API key authentication

🏢 Multi‑Tenancy & Cost

• Isolated vector namespaces per tenant
• Redis‑backed rate limits + monthly quotas
• Per‑query cost tracking with Prometheus

📊 Observability

• Full OpenTelemetry tracing across pipeline stages
• 15 Prometheus metrics (latency, cost, hallucination score, citation coverage, etc.)
• Auto‑provisioned Grafana dashboards
• Structured JSON logging with OTel trace/span injection

⚙️ API, SDK & CLI

• FastAPI REST API (ingest, query, tenants, feedback)
• Thin Python SDK (async + sync)
• Typer CLI with ingest, query, evaluate, serve, health

📈 Evaluation

• RAGAS evaluator for faithfulness, relevancy, precision
• Numeric accuracy scorer (LLM‑as‑judge)
• Expanded golden dataset: 22 financial filings
• CI regression gate (faithfulness drop >5% blocks PR)

🏗️ Infrastructure

• Hardened multi‑stage Dockerfile (non‑root)
• Compose stack: Redis, OTel, Prometheus, Grafana, Jaeger
• Kubernetes manifests: Deployment, Service, HPA, NetworkPolicy, PDB, ConfigMap, Secrets, PVC
• Helm chart + Kustomize overlays (dev/prod)
• Terraform stubs for EKS, RDS, ElastiCache, S3, IAM

🧪 Testing

• 322 test functions across 26 files
• Unit, integration, property‑based, API integration, chaos engineering
• k6 load test: p99 < 8s, <1% error rate

📚 Documentation

• 43 Markdown files with full MkDocs Material site
• 7 ADRs, tutorials, API/CLI references
• Troubleshooting + performance/cost tuning guides
• Streamlit demo UI

---

✅ Release Notes Summary

• Zero syntax errors across 180 files
• Enterprise‑grade multimodal RAG pipeline
• Production‑ready infra, observability, and evaluation gates

---