Changelog
Release Notes
Track the latest updates, new features, and improvements to the PIM AI Global platform.
v1.3.0April 2026Feature
Caching, Observability & Polish
- Embedding cache: repeated queries skip the OpenAI embedding API call (30-day TTL, Supabase-backed)
- Response cache: identical query+collection+model combinations return instant cached answers (24-hour TTL)
- X-Cache HIT/MISS headers on all query responses for transparency
- Structured query logging with timing breakdown (embedding, retrieval, LLM generation)
- Analytics dashboard at /analytics with latency charts, cache hit rates, and token usage
- Copy-to-clipboard button on AI-generated answers
- Query history stored in localStorage with one-click recall
- Row-Level Security (RLS) enabled on all Supabase tables
- Environment variable validation via Zod at startup
v1.2.0April 2026Feature
UX Quick Wins & Security Hardening
- Markdown rendering for AI answers — headings, lists, bold, tables, code blocks display correctly
- Keyboard shortcut: Cmd/Ctrl+Enter submits query from the textarea
- SEO metadata: Open Graph tags, Twitter cards, and per-page titles across all pages
- Security headers: X-Content-Type-Options, X-Frame-Options, Referrer-Policy, Permissions-Policy
- Accessibility improvements: ARIA labels on form controls, aria-live on streaming content, aria-expanded on mobile nav
- Updated tech specs and About page to reflect multi-LLM support
v1.1.0April 2026Feature
Performance Optimization & Multi-LLM
- Multi-LLM router: choose between Claude Sonnet, Claude Haiku, Claude Opus, GPT-4o, GPT-4o Mini, and o3-mini
- Response streaming via Server-Sent Events (SSE) — answers appear in real-time
- Rate limiting: 20 requests per minute per IP with Retry-After headers
- Stats endpoint optimization: 12 queries reduced to 3 parallel RPCs with 60s caching
- Code quality: runtime type validation, proper error handling, unique constraints on chunk tables
- Removed unused browser Supabase client, fixed page number calculation in ingestion
v1.0.0April 2026Release
PIM AI Global — Initial Platform Launch
- Launched PIM AI Global RAG platform with Next.js 16, Tailwind CSS, and TypeScript
- Three curated vector databases: Global PIM Good Practices, IMF PIMA Reports, and World Bank Public Finance Reviews
- Semantic search powered by OpenAI text-embedding-3-small (1,536 dimensions) with Supabase pgvector
- AI-powered answer generation with source citation
- Multi-format document ingestion pipeline supporting PDF, DOCX, and TXT files
- Real-time collection statistics dashboard with technical specifications
- Query interface with collection selector dropdown and similarity-ranked source display
- PIM PAM branding with official logo in header and footer
- Responsive design with dark navy header/footer matching pim-pam.ai styling
- Deployed on Vercel with environment-based configuration
v0.1.0April 2026Feature
Document Ingestion & Vector Database Setup
- Ingested 12 Global PIM Good Practices documents (~8,111 text chunks)
- Ingested 107 IMF PIMA Reports covering 30+ countries from 2016–2025 (~13,886 chunks)
- Ingested 316 World Bank Public Finance Reviews spanning 2015–2026 (~161,884 chunks)
- Supabase pgvector schema with IVFFlat indexing (100 lists) for fast cosine similarity search
- RPC functions for similarity matching and collection statistics aggregation
- Intelligent text chunking (1,000 chars with 200-char overlap) for optimal retrieval